PH热榜 | 2026-05-20

@emma_watson21 Thank you~

@emma_watson21 Thank you so much.

Would definitely encourage you to try it out and see how it fits your workflow, always happy to hear feedback from early users.

@emma_watson21 Thank you very much.

@itsluo Sure,but you’ll need to provide specific prompts when chatting with the AI, so it can carry out actions such as pricing updates, inventory adjustments, or marketing campaign changes according to your requirements.Welcome to reach out to us via  customer.support@storeclaw.ai.

@itsluo 

One thing I'm curious about: can sellers set approval layers before actions run? Especially for pricing, inventory, or campaign changes.

@itsluo Yes — approval layers are core to the design. Operators set scope per action type: lower-risk work (content, listing health, lifecycle) can run autonomously, while pricing, inventory, and budget changes queue for review by default. You dial it up or down per category as trust builds. Higher blast radius = more conservative default. Nobody wants an agent cutting prices autonomously at 2am.

@peng_wood Thanks Wood, that's the framing we've been working toward, most "AI for ecommerce" still ships as a content generator with extra steps. Means a lot that it reads operations-first to you.

@peng_wood Thank you,we’ve long felt the space was too focused on content generation, so we built StoreClaw to be operation-first. It’s great to hear that resonates with you.

@peng_wood 

Thanks, Wood， that’s exactly the direction we’ve been pushing for. A lot of “AI for ecommerce” still ends up being content tooling with extra steps, so it’s good to hear the operations-first angle is coming through.

@surabhi_minocha That’s a great perspective, and we completely agree. The question is difficult to measure precisely,but one thing we’re certain about is that we’re continuously working toward that goal.

@surabhi_minocha 

Exactly, agents handle high-frequency execution, while operator judgment is encoded as guardrails and constraints.

The operator shifts from micro-decisions to defining and refining the boundaries the system runs.

@surabhi_minocha You've named the actual hard problem. Skills are commoditizing — anyone can wire up an SEO or copy agent. Judgment is the moat: knowing when not to optimize. Our approach is to make operator judgment first-class context that travels with every skill run — seasonality, inventory posture, margin tolerance, brand guardrails, launch phase. The engine isn't optimizing in a vacuum. Day-one match for a 15-year operator's gut? No. But the gap closes faster than people expect once judgment lives in the system instead of one person's head.

@jaredl Really appreciate this perspective.

That’s exactly why we built the system around controlled autonomy, agents can act, but never override judgment. Brand consistency and merchant intent stay at the center of every action.

Would love for you to give it a try when you’re ready and see how it behaves in a real store context.

@rocsheh 

Thanks a lot for the thoughtful feedback, Zepeng She， especially the channel-by-channel diff preview idea, that’s very aligned with how we think about multi-platform safety and control. Cross-channel normalization is one of the trickiest parts, so making changes transparent per platform is definitely something we’re exploring.

On differentiated goals: not fully exposed as a product setting yet, but the underlying direction is exactly that different optimization strategies based on product intent (clearance vs. hero SKU vs. new launch). It’s on the roadmap as we expand from “single-plan optimization” to “portfolio-level control.”

Really appreciate you digging into the practical edge cases, that’s exactly where the system needs to be strong.

@rocsheh Thank you so much for the sincere recognition and professional advice!We’ll keep polishing the product to fit diverse operational needs, really grateful for your great support.

@rocsheh Really appreciate such a thoughtful take. We’re focused on giving merchants real execution power across channels, while keeping control and clarity at the center of every action. Excited to keep evolving this with feedback like yours and push the boundaries of what commerce agents can do.

Would love for you to try it out and see it in action on your own setup.

@ray_luan Thanks Ray, appreciate that especially coming from a fellow builder. Crowded space is right; we'll keep our heads down and try to earn it.

@ray_luan Much appreciated!

@ray_luan 

Thanks Ray — that means a lot, especially from someone building in the same space. It’s definitely a crowded field, so we’ll just stay focused and keep shipping until it speaks for itself.

@ytubviral We adopt multi-model collaboration to gather advantages from each one, and we are continuously optimizing scheduling logic to lower latency and ensure smooth usage experience.

Our approval mechanism supports fine-grained permission settings. Merchants can set personalized rules freely. Operations with small impact can run automatically, while relatively important adjustments will be paused and wait for your manual review. We keep optimizing this function to fit various usage scenarios.

If you require more detailed technical consultation, please write directly to: Customer.support@storeclaw.ai

@ytubviral We match different models reasonably to make full use of their respective strengths, and we keep tuning our system operation process to cut down waiting time and improve overall response efficiency.

Fully understand this concern, and our platform supports flexible and detailed permission management. Sellers can make exclusive rules by themselves, letting low-impact adjustments run automatically, and all major changes will be held back to wait for your confirmation first. We keep upgrading this practical feature all the time.

And more information can be inquired from: Customer.support@storeclaw.ai, we are glad to hear from you~

@ytubviral Thanks Javier — these are sharper questions than we can do justice to in a comment thread, and you're clearly the kind of operator we'd want to go deep with. Drop us a line at customer.support@storeclaw.ai and we'll set up a real conversation on the architecture and approval-control side. Would genuinely like to compare notes with another agent-builder.

@nina563 Thanks a lot for your warm wishes and support! We’re delighted to build this all-in-one solution covering Amazon, Shopify and DTC stores. We will keep polishing cross-platform functions to help merchants manage multi-channel businesses more effortlessly.

@nina563 Thanks Dan — appreciate the support from the Genstore team.

@artstavenka1 We totally understand your concerns about Amazon compliance and safety! StoreClaw integrates exclusively via Amazon’s official SP‑API, fully adhering to their rate limits and TOS for all automated listing and pricing activities.

For more details on integration and security, feel free to contact our support at custosmer.support@storeclaw.ai.

@boyuan_deng1 Storeclaw is exactly what you need. It’s designed to significantly improve efficiency and productivity for small teams like yours. Welcome to subscribe and give it a try!

@boyuan_deng1 

Definitely!

@boyuan_deng1 That's the whole bet. "Monitor the AI" is just another task — we built it to close its own loops so you don't inherit a new job. 2-person teams are exactly who this is for. DM me if you want early access.

@abod_rehman You can share your industry background and historical business data with the AI, and it will identify patterns based on your business performance and industry cycles.

That said, providing clear context and goals in your prompts can help the AI make even more accurate decisions for your business.

@abod_rehman Fair concern — "panic and cut everything" is the failure mode of most optimization tools. Two things. One, the engine reads category and competitive seasonality, not just your store in isolation — so day one in January, it already knows February is a trough for your category. Two, your own store history gets ingested at connection, so your specific seasonality shape (Q4 ramp, summer dip) feeds into how current performance is read. Slow February isn't a problem to solve — it's a baseline to plan against.

@abod_rehman 

No worries at all — we’ve got that covered. StoreClaw is built to handle those shifts safely without overreacting 🚀

@alexia_li 

What separates this from tools that only adding AI features is the action layer. Every existing tool is still a dashboard — it shows you what's happening and tells you what to do. StoreClaw actually doing the thing is the leap. 

@alexia_li 

That’s exactly the gap we’re targeting. StoreClaw isn’t built to advise operators — it’s built to be one, with real execution across workflows instead of static insights.

@alexia_li Nailed it. Dashboard + chatbot is still a productivity tool — the human is doing the work. An engine that executes changes the category, and the hiring math with it. The question shifts from "do I need another SEO/lifecycle/content hire" to "what channel can I launch next without adding payroll."

@eeeeeach 

Really appreciate that! We built StoreClaw to move beyond “AI suggestions” and actually handle execution for sellers — from product sourcing to store ops and optimization.

This launch is just the beginning, but we’re excited to finally show what autonomous commerce can look like at real scale. Looking forward to sharing more wins (and lessons) as sellers start pushing it in production 🚀

@eeeeeach Appreciate it! The demo version of autonomous is easy — the at-scale version is where most tools quietly break. That's the bar we're holding ourselves to.

@eeeeeach 

I've used Shopify Magic and Amazon's own AI tools. They're both surface-level — autocomplete for listings, basic analytics framing. StoreClaw running actual commerce decisions in the background is a completely different scope of ambition.

@michelle_thirteen Really appreciate your recognition! It’s our core goal to solve real operational pain points.

@michelle_thirteen Many thanks for the sincere wishes! We keep focusing on easing operational troubles for merchants. Welcome to join us and experience the product.

@michelle_thirteen Thanks, Michelle. We’ve always believed the real unlock in commerce comes from reducing operational friction and helping founders move faster with confidence.

@wenyi_yu Truly appreciate your support and good wishes! We’re glad we can finally bring this solution to all sellers.

@wenyi_yu Really appreciate this.

We’re building this for exactly that, giving e-com operators something they can actually run with, not just experiment on. Would love for you to try it when you’re ready.

@wenyi_yu Thanks so much for your warm support and best wishes! We truly appreciate it. We will keep optimizing the product to better serve all sellers. You are warmly welcome to sign up and experience it.

@leah_leee Thanks for saying that,it’s exactly why we built this. The e-commerce space is flooded with AI tools that only generate content, so we went all-in on operations-first automation. Glad that resonates with you.

@leah_leee Really appreciate that.

We’re focused on building something that actually helps merchants execute, not just generate ideas. Excited to keep pushing this forward together and see how far we can take it with the community.

Would love for you to try it out when you’re ready, we’re here if you need anything.

@leah_leee Many thanks for your kind words! Moving beyond mere content generation to serve actual operational demands is our original intention.

@satomi_yu  @hunter_powabase Thanks Hunter, both good questions.

On the data threshold: there's no hard minimum, but the agent honestly degrades what it claims based on what your data supports. On a low-traffic store, it'll lean more on category-level benchmarks, competitor signals, and structural audits (does your schema work, are your titles missing intent keywords, etc.) — and less on conversion or attribution claims that need real volume to be statistically real. When it does surface a recommendation that's data-thin, it flags the confidence level rather than presenting it as solid. The principle we hold: never pretend a recommendation is data-backed when it isn't.

On Etsy: it's labeled "Coming Soon" on our connectors page today. Etsy operators are exactly who we'd want to hear from while we're building it out, so if that's your store, drop us a line.

@hunter_powabase 

No hard data minimum — the agent adjusts confidence to the quality of available data and won’t overclaim insights it can’t support.

@hunter_powabase Great question,this is exactly the scenario we designed StoreClaw for.There is no minimum sales data requirement to get started. It will never push high-risk changes, and every action requires your manual approval before going live.

@tripplep Thanks Prajesh, that genuinely means a lot — especially on day one. If anything surprises you as the agent gets deeper into your store, good or bad, I want to hear it.

@tripplep Thanks Prajesh, this is the exact loop we built it for, and seeing someone actually run it on day one is the best news I've had today. Reach out anytime, especially if something doesn't behave the way you'd expect.

@tripplep Thanks so much for the feedback and for sticking with us.

@luz_bidelspach Pretty much all AI-driven tasks on StoreClaw use credits, like writing product content, optimizing listings, doing competitor research and making marketing copies.

Actually there’s no set credit cost for any single feature. It all depends on how heavy the work is, like how long the content is or how many products you need to analyze. That’s why we don’t list fixed rates upfront.

No worries though, you can always check the exact credits used in Usage Records under Settings right after each task finishes.

@luz_bidelspach Task-generation actions started in StoreClaw generally consume credits, including AI-powered tasks such as content generation, listing optimization, competitor analysis, and campaign copy generation.

The number of credits consumed by a single task is not fixed. It varies based on output complexity, such as content length or the number of items being analyzed. For that reason, we do not publish a fixed per-task price list in advance.

After each task is completed, you can review the exact deduction in Settings -> Usage under Usage Record.

@luz_bidelspach Thanks Luz — to actually answer your question rather than around it: we don't have a clean ceiling number to quote yet because the beta is small and the heaviest jobs (full-catalog audits on large stores) haven't been stress-tested. Directionally, single-product actions are tens of credits, mid-sized jobs are low hundreds, and the biggest jobs we've seen are in the low thousands. We'll publish a real ceiling once we have meaningful data across more store sizes.

@elijah_smith6 Thanks Elijah, good catch on the order of operations. Yes — duplicate meta descriptions are surfaced in the audit step as their own flagged issue, alongside missing descriptions, length violations, and keyword cannibalization across pages. The agent only proposes rewrites once you've seen the diagnostic, and you approve which ones to actually touch.

@elijah_smith6 Absolutely it does! Our SEO audit will first spot and mark out all duplicate meta descriptions clearly ahead of time. This way you can know all existing issues upfront, and the AI will only start rewriting after sorting these problems out properly.

@elijah_smith6 Sure thing! Our SEO audit can easily pick out all repeated meta descriptions in advance. You can get a clear overview of these issues first, then let the AI carry out targeted rewriting work later on.

@natalia_iankovych Haha, what a coincidence! We’ve always focused on building a data-driven optimization experience. Right now, we support integration with major marketing tools like Google Analytics, syncing key data for analysis and recommendations. We’re also continuously expanding compatibility with more popular marketing tools, strengthening data connectivity to deliver more comprehensive, precise optimization plans.

@natalia_iankovych Thanks Natalia, and that's a fun coincidence — please tell your friend hi.

To add to what Satomi said: Google Analytics is live today (GA4, with traffic, attribution, conversions, and realtime). Google Ads is also live — ROAS, ad spend, campaign optimization. Google Search Console is on the roadmap, not shipped yet. It matters because GSC tells us what actually happened in search after the agent ships a change — without it, we're reading one side of the conversation.

Full current list of connectors is at storeclaw.ai/app/connectors. The priority order for what gets built next is shaped by which tools beta operators tell us they're already in every day. If yours isn't on the list yet, that's exactly the kind of feedback we'd want to hear.

@natalia_iankovych It feels great that our product matches your idea so well. We are working on supporting integrations with Google Analytics and more popular marketing tools. We keep optimizing data connection features, so we can gather sufficient data to provide more practical optimization advice for everyone.

@krishna_tyagi5 Thank you so much for your kind praise! We highly value transparent decision logic display, letting users clearly know the thinking behind every suggestion. We will keep polishing details and optimizing user experience continuously.

@krishna_tyagi5 Thanks Krishna — that "show its thinking" piece is the one I personally care about most. Glad it landed.

@joshua_martinez7 Yes, the 300 free credits do include the backlink generation skill—no premium upgrade required. It counts toward general credit usage, just like SEO fixes, Listing inspections, page optimizations, and other core features. For details on exact credit consumption, feel free to contact our support team at custosmer.support@storeclaw.ai.

@joshua_martinez7 Feel free to try it on your store whenever you’re ready, we’d love to see how it performs for you. My team is always on standby and happy to help if you need anything.

@emily_carter18 StoreClaw uses a flexible inherit + independent customization approach: Variants inherit the base meta description from parent listings by default, and you can generate and edit unique meta descriptions for each variant separately (by color, size, etc.) for better SEO and platform compliance (Amazon, Shopify, etc.).

For detailed configuration support, feel free to contact us at custosmer.support@storeclaw.ai.

@emily_carter18 Thanks Emily, and Satomi already covered the mechanism well. To add the operator side: inherited is usually the right default. Per-variant meta descriptions only earn their keep when variants compete for distinct search intent — e.g., "red leather wallet" vs. "black leather wallet" where the color is a high-volume search term, or sizing where buyers are explicitly searching by size.

When variants are just stocking units of the same product (same intent, same buyer), per-variant overrides usually create maintenance overhead without ranking lift. The agent flags which of your variants meet the "distinct intent" bar before recommending overrides — so you're not editing 47 color variants just because you can.

@joshua_cooper2 Couldn’t agree more — execution is everything. StoreClaw’s action plans are data-backed, high-revenue impact, and low-effort to implement, so the user rejection rate is very low. We keep refining our logic based on real user feedback to make every suggestion actionable, effective, and worth your time.

@joshua_cooper2 Thanks Joshua, and you're asking the right question. Honestly: we don't have a clean acceptance-rate number yet that I'd trust enough to quote. The beta cohort is still small and the recommendations we ship today are heavily weighted toward the low-effort, high-confidence end of the action space — which biases the number upward.

The signal I actually watch is which categories of recommendations get rejected. Pricing changes get rejected most often (operators want to call those themselves). Listing copy and schema rewrites get rejected least often. That tells us where the agent is genuinely useful vs. where it should be quieter.

Once we have meaningful volume across more diverse stores, I'd rather publish a real number than estimate one now.

@cruise_chen Thanks so much for the congratulations! StoreClaw offers ready-to-use intelligent templates today, while it continuously learns brand-specific patterns and data behaviors to adapt to your store style, Listing habits, pricing strategies and operational preferences over time. Feel free to try it out!

@cruise_chen Thanks. Honestly, both — by design. Skills give the engine a strong baseline so it's productive day one, no "wait six weeks to learn your brand" tax. Then brand-specific patterns layer in over time: voice, what's worked in your category, approval patterns, which playbooks have hit. Template-driven day one, increasingly brand-specific from there. What we deliberately avoid is making operators wait for a learning curve before the engine earns its keep.

@james_carter35 Couldn’t agree more. We focus heavily on practical and actionable features. The system keeps full operation logs, and most regular adjustments can be undone easily. We are also constantly improving relevant rollback functions to bring users more flexible and reliable experience.

@james_carter35 Thanks James, and to add a layer to what Lena and Satomi already said: every change StoreClaw ships gets logged with a timestamp, a diff, and a one-click reversal. Most edits — listing copy, schema, descriptions, pricing — roll back cleanly in seconds.

Where it gets messier is changes that compound (a price drop that's already converted 30 orders, an inventory move that's already shipped). For those, the agent flags "irreversible from this point" before asking you to approve. We'd rather pause and explain than ship something you can't take back.

@antonio_manuel1 Thanks a lot! The generated product copy supports full customization. You can set brand tone, writing style and target audience freely, and adjust content length and marketing focus to perfectly match your store style and different platform demands.

@antonio_manuel1 Thanks Antonio, and to add to what Satomi listed: customization works on two layers.

First, a global voice profile built from your existing content — tone, vocabulary, sentence rhythm, words you avoid. The agent matches new copy to that profile by default, so most of the customization happens passively.

Second, explicit overrides per product line or category. If your athletic SKUs need a punchier voice than your loungewear, you set that once and the agent generates accordingly. Same for length, platform-specific format (Amazon bullets vs. Shopify hero copy), and which value props lead.

@thamibenjelloun Hi Thami, congrats on the launch. Do you just surface issues or can your agents resolve some of the common ones?

@thamibenjelloun 🏆

@thamibenjelloun Emailing made me meet great people and build an amazing team!

@garrytan  @othman_katim  @karimbenkeroum  @manal_essalek1  @naimz  @daniel_nwankwo  @bengeekly and all the others. Proud of this team.

mailX is really the result of different strengths coming together: deep deliverability experience, product thinking, engineering, AI workflows, customer support, and a lot of real problems we saw through Mailwarm over the years.

That’s what makes this launch special for me.

Not just a tool, but years of learning turned into something simple for humans and AI agents.

Let’s go team 🚀

@bengeekly  heavy on the "boring stuff with scary names" lol. It's been lovely working with you and everyone too. To the launch and beyond 🚀

@busmark_w_nika Definitely something to keep in mind for our Roadmap 😉

Hi @busmark_w_nika, thank you for raising this!

That's the question we keep asking ourselves. Today, we surface what's broken and guide the fix, full automation is the logical next step. Which part of the setup would you most want automated first?

@busmark_w_nika No. Our focus is not to create the email itself and send it.

Many tools already help with copy, personalization, and sequences.

Our focus is the part after that: making sure the email has the best chance to land in the inbox.

So mailX diagnoses what is wrong, through the web app or through AI agents via API & MCP.

The future for us is more about helping agents check, fix, and monitor deliverability safely, not replacing the creative part.

@othman_katim  @zaczuo That’s exactly what we’re focused on with MailX: keeping the system transparent so you always understand why something is happening, not just what happened.

Really appreciate the support 🙏

@zaczuo thanks!

The agent layer thesis is real, and the piece nobody talks about yet is the plumbing underneath. An agent that lands in spam is worse than one that didn't send at all, because the user assumes the action happened.

That's the wedge we keep coming back to. Are you building on top of the email layer?

@othman_katim  @zaczuo Thank you!! Really thoughtful perspective. Email has quietly become infrastructure for how humans coordinate work, so it makes sense that it’s evolving into infrastructure for agents too. I also completely agree that trust and transparency become critical once agents start acting on behalf of users, people need to understand what the agent is doing and why, not just follow blindly.

@manal_essalek1  @byalexai Yes at mailX, we are focused on turning issues into actionable fixes rather than just showing raw data. So yes, the idea is to guide you with clear next steps when something is wrong, and give you the right tools to solve deliverability problems. :)

@byalexai thank you!

Yes, and that’s actually one of the most valuable parts of it. It doesn’t just flag issues, it also gives simple, actionable recommendations so you know what to fix without digging too deep.

@byalexai  Would love to see ZeroHuman and MailX interacting !

@yurimhln That's correct! we pay close attention to differences between email clients, especially Gmail since they’ve become much stricter lately. We optimize for rendering and deliverability across major clients (Gmail, Outlook, Apple Mail, etc.) because each behaves differently :)

@yurimhln Gmail is evolving their way of managing email screening, this will impact everyone email KPIs, like open rate.

@yurimhln Yes. Deliverability is not the same across providers.

Gmail, Outlook, Yahoo, Microsoft 365… each one has its own rules and signals. Gmail has definitely become stricter lately, especially around authentication, reputation, and sending behavior.

That’s why we check issues at the domain/setup level and are moving more and more toward provider-specific diagnostics.

@musharofchy Thanks

@dawid_baranowski Appreciate it 🙌 That’s exactly one of the gaps we wanted to solve. A lot of these checks still require jumping between multiple tools and manual debugging steps.

Would really love to hear your feedback once you try it especially from someone already doing this stuff manually :)

@dawid_baranowski Would be nice to have your feedback after you test it !

@dawid_baranowski Love this. The raccoon copy is great. Your founders' entire fundraise depends on those emails landing in a VC inbox instead of the promotions tab, and a silent spam send is worse than a no because there's nothing for the founder to act on. Would love your feedback once you spin it up.

@surabhi_minocha We’re already moving in that direction in all transparency. I think inbox providers will increasingly score behavioural patterns, not just technical authentication.

As AI-generated outreach scales, signals like engagement quality, sending behaviour, personalization depth, and human-like interactions will probably matter even more than they do today.

@surabhi_minocha I think they already look at behaviour patterns like engagement, complaint signals, and sending consistency, even if they don’t call it “AI patterns” yet. If AI content makes inboxes noisier, they’ll likely just tighten how they score engagement quality, not the content type itself.

@surabhi_minocha Yes, and I think the issue won’t be “AI content” itself. It will be AI-scaled bad behavior: same patterns, same timing, same low relevance, too much volume. Inbox providers will punish that fast.

Are you already seeing this with AI-generated outbound?

@konstantimb exactly that :))

@konstantimb To the Moon!

@konstantimb THANK YOUUUUUU

 Love this perspective @manal_essalek1 . This is exactly why we built mailX: email deliverability should not feel like a black box reserved for technical people. Most teams just want to know what is broken and what to fix.

Proud to see the team turning years of customer pain into something simple, clear, and useful 🚀

@asheer_ahmad Exactly. Founders don’t need more raw metrics. They need to know what is broken, why it matters, and what to fix first.

@asheer_ahmad Exactly! people can see SPF/DKIM/DMARC records or deliverability metrics, but translating that into what’s actually broken and what should I do next? is the hard part.

That’s the gap we’re trying to close with mailX :))

@asheer_ahmad Got it right!!!

@natalia_iankovych Honestly the is to many things to handle with deliverability.
Checkers

• SPF

• DKIM

• DMARC

• BIMI

• SMTP Checker

• IMAP Checker

• Blacklist Checker

Generators

• DMARC Generator

• SPF Generator

• BIMI Host

Finders

• SMTP Finder

• IMAP Finder

DNS

• MX Lookup

• TXT Lookup

• CNAME Lookup

• PTR Lookup

• DNS Lookup

Here is what we can check right now if you want to compare with anything else, and the main difference is instead of understanding this keywords made for machines, we made it accessible fe AI agent check them via MCP.

@natalia_iankovych Fair question tools like Mail Tester are great for quick checks and basic validation.

With mailX, the goal is less about just giving a score, and more about helping users actually understand what’s wrong, why it matters, and how to fix it especially over time and across different workflows.

And yes, Google and Microsoft can behave very differently deliverability-wise, so provider-specific visibility is definitely something we care about.

@natalia_iankovych Mail-tester is useful. The difference we’re building toward with mailX is broader: not only a one-time score, but a full diagnostic layer for web, API, and AI agents through MCP. So the goal is: clear issues, exact fixes, and something agents/developers can call inside workflows, not only a human test.

On Google vs Microsoft: today we mainly diagnose the setup and risks through mailX. Provider-specific inbox placement is part of where we’re going, because Gmail and Microsoft don’t behave the same at all.

But on mailwarm we do provide a SPAM Score per email provider. @manal_essalek1 will be happy to help you on that.

@artstavenka1 Of course, actually the DNS protocols are free to setup, and highly recommended to everyone by Gmail, Microsoft and all Email providers

@artstavenka1 Definitely! In some ways, low-volume founders are even more sensitive to deliverability issues because every email matters.

If you’re sending founder outreach, partnerships, recruiting emails, or early sales conversations, losing even a handful of emails to spam can have a real impact.

@artstavenka1 Yes, definitely. Low volume is actually the best moment to catch issues before they become expensive. If your setup, authentication, or reputation is weak, scaling only makes the problem bigger.

@german_merlo1 Exactly, email is one of those fields that looks simple from outside, but becomes very deep when you start scaling. Appreciate the support!

@german_merlo1 Thanks for the support

@german_merlo1 Thank you so much. Email looks simple until you start scaling. That’s why we’re trying to make deliverability easier to understand and fix. Appreciate the support!

@vikramp7470 Thank you !! And yes, that’s part of the broader problem we’re trying to solve. Authentication is foundational, but deliverability is also influenced by sending behavior and content quality.

@vikramp7470 Thank you! This is not exactly what themailx itself does but it’s extension MailWarm does. Both tools work hand in hand!

@vikramp7470 Thanks!

On copy scoring: yes, Mailwarm has a spam checker that scores the email content before you send and flags the specific trigger words and patterns hurting placement. It's per-email scoring, you paste in the draft and get a verdict with the flagged issues. https://www.mailwarm.com/spam-checker

@theokitsberg Really good question. Maybe agents will eventually need deliverability-awareness built into their workflows, because otherwise they’ll just optimize for volume and unintentionally damage sender reputation over time. The goal with mailX is helping agents monitor domain health, detect issues early, and adapt instead of sending blindly.But yeah I appreciate your question :))

@theokitsberg Good question. In practice the agent calls the MailX MCP before it sends, basically as a pre-flight check. It pulls SPF, DKIM, DMARC, blacklist status, MX for the sending domain, and gets back a verdict it can act on: send, hold, or fix first.

If something's broken (say DMARC is set to none), the agent can call the SPF/DMARC generator tool to produce the corrected record, hand it back to the user, then re-check on the next run.

The shift is from blind 'spray and pray' to an agent that knows before it sends whether the email has a chance of landing.

@theokitsberg In practice, mailX is the diagnostic layer the agent calls before sending or scaling.

It can check the setup, spot risks, and tell the agent: “safe to continue” or “fix this first.”

That’s the direction we believe AI outbound needs.

@abhishekr_ai You'd defo want some level of domain warm-up first. Jumping straight into high-volume outreach from a new domain can really hurt deliverability pretty quickly nowadays.

mailX is designed to help users monitor those signals and scale sending more safely over time but we have mailwarm designed for domain warmup :))

@abhishekr_ai I’d first check the setup with mailX, then warm up gradually with Mailwarm if the domain is new or risky. High volume from day one is rarely safe.

Is it a fresh domain or one already used for sending?

@abhishekr_ai You still need some form of warm-up, there’s no truly “safe” way to hit high volume from day one on a fresh domain.

@basheer_rjoub That’s exactly the kind of workflow mailX is for 🙌
When your inbox volume gets that high, even saving a few seconds per email adds up fast.

@basheer_rjoub Thanks God, AI agents will save us from inbox anxiety!

@basheer_rjoub That has always been our goal 😉

@robert_hughes5 For sure, I think a lot of infrastructure and developer tools will need to become more agent-friendly over time, not just human-dashboard-friendly.

@robert_hughes5 Exactly. Agents will need tools they can actually call, not just dashboards humans read. That’s the shift we’re building for with API + MCP.

@robert_hughes5 Exactly. It feels like the shift now is from “AI-assisted tools” to tools that are built assuming agents will actively interact with them.

@enviouscoder What a good read! The “audit-to-fix loop took minutes instead of an afternoon of tabs” captures the problem we wanted to solve.

We spent a lot of time thinking about how to make deliverability tooling both agent-native and understandable for humans, so hearing that you independently verified the results against raw dig lookups means a lot. Really appreciate you taking the time to test it this deeply on day one :)))

@enviouscoder That’s the kind of use case we built mailX for. Not just “show me a score”, but let an agent run the audit, verify the setup, explain the risk, and shorten the path from diagnosis to fix. Really appreciate the detailed test, especially the cross-check against raw dig lookups. Accuracy is critical when DNS changes are involved.

@enviouscoder This was honestly amazing to read 😄

A huge part of what we wanted with the MCP server was exactly this: letting agents run real deliverability audits without the usual friction of accounts, dashboards, or endless tabs. Really glad the results held up against manual dig verification too.

@andrzej_zarod good question. If you only send a small number of important emails like investor outreach, sales conversations, partnerships, support replies, etc missing the inbox hits harder. Higher volume makes the patterns more visible, but the underlying problems affect everyone.

@andrzej_zarod Low-volume are definitely benefiting setting DNS properly should be done by everyone. It's the highest results for the lowest effort/cost (Free)

@andrzej_zarod Yes, useful even at low volume. Actually, it’s better to catch deliverability issues before scaling.

At high volume, the pain is bigger. At low volume, mailX helps you avoid damaging the domain early. It the first step before starting email warmup.

@munis_abbas Oh that sucks! see that’s exactly the kind of deliverability issue we kept seeing over and over.

Everything looks fine, campaigns are sending normally aaand then a chunk of emails quietly starts landing in spam. Glad the timing of mailX resonates :)))

@munis_abbas Here is the best steps to follow: Install our MCP in your Claude account, then ask it what you need to fix. If you need anything more, book a call with @othman_katim

@munis_abbas , I will be glad to personally help you resolve this, let's connect on Linkedin

@francesco2689 This is an interesting parallel. We’re starting to see more conversations around visibility, trust, and observability for agent actions especially as agents become more autonomous. Once agents are acting on behalf of users, people want to understand the what and the why and whether the outcome was actually successful.

@francesco2689 Thanks! Glad it landed.

Same failure mode you described: the agent acted, the user trusted, the brand never saw. Visible to one side, invisible to the other.

Yes, teams ask for it, but framed as inbox placement so far. The agent-recommendation version isn't on most radars yet. It will be the moment founders realize their AI citation rate can drop silently the same way inbox placement always did.

What are you building?

@francesco2689 Yeah, this is exactly the direction things are heading.

There’s a growing gap between “what the system did” and “what the brand knows happened.” With email it’s spam filters and deliverability black boxes. With AI agents, it’s recommendations, tool calls, and decisions happening with zero audit trail for the brand. We’re starting to see teams ask for that visibility, not just logs, but explanations of outcomes: why something was recommended, what alternatives were considered, and when a competitor gets surfaced instead.

@saad_el_gueddari That’s exactly the problem we kept seeing 😅 Once you manage multiple warmed domains, half the battle is just figuring out where the issue actually comes from.

And yes the MCP can monitor deliverability continuously across multiple sending domains. The goal for us is to help agents (and humans) spot issues early instead of debugging after performance drops.

and Thanks a bunch :))

@saad_el_gueddari You nailed the diagnostic problem. Auth, reputation, workspace, and content all fail differently, and you can't fix what you can't locate.

On the MCP: today, it's point-in-time per domain. You can run at the frequency you would like to. Continuous monitoring and enrichment by external data from your email platform is on our radar. So many things are coming!

@saad_el_gueddari Exactly, finding where the problem lives is usually half the work.

Today, MCP is point-in-time per domain, but an agent can loop through multiple sending domains and run checks. Continuous monitoring is the next layer. That’s where we’re going with mailX + MailAdept.

@campritchard mailX is not something we built because “AI agents” are trending. It comes from years of seeing the same email deliverability pain inside real teams. The goal was to turn that experience into something simple, fast, and usable by both humans and agents.

@arnestrickmann been running into the electron weight tradeoff myself — solo dev, three or four dev tools open at once, and claude code in plain iterm is by far the lightest thing in the lineup. running four parallel agents on long jobs, what matters is whether emdash sits closer to a stack of terminal tabs or to a heavy editor like WebStorm — that's the gap that decides whether you can keep it open all day.

@arnestrickmann How is Emdash thinking about orchestration reliability as teams scale the number of concurrent agents?

@arnestrickmann Hi Arne, big congrats on the launch 🎉

Parallel agents with isolated worktrees is the part that catches my eye. Most multi agent desktop tools today are basically tab managers, you flip between sessions but they share the same checkout. Isolated worktrees actually let agents work on the same repo without stepping on each other.

Question on the merge story. When two agents finish parallel tasks that touch overlapping files, what happens? Does Emdash surface the conflict before opening PRs, auto rebase one, or just ship both PRs and let me sort it out on GitHub?

Running a few coding agents in prod today and this is the exact wall I hit, will give Emdash a try this week.

@dominikkoch that's amazing! And exactly what we initially built Emdash for, to make the wrangling with worktrees and terminal windows a little bit less annoying.

And yes, we'll ship a chatUI similar to T3code soon! We started out with CLIs first because this allows us to ship fastest and offer the full functionality of the CLIs without having to play continuous catchup as they expand their feature set with features such as /btw in claude code.

@lakshminath_dondeti great question! And no, we natively embed the CLI so in the eyes of Claude Code / Anthropic it is as though you're opening Claude Code in any other Terminal emulator like Ghossty, iTerm, etc..

@elliot_grant1 exactly! That's what we've been thinking. There's a lot of convergence in this space anyways and we want to build the foundation and primitives for how everyone works with (coding) agents

@maxjbre very happy you like it!

@gabor_hollbeck 🙏 thank you for all your invaluable user feedback

@alexanderfarr thank you, Alex!! And congrats on your launch last week. Send all feedback my way!!

@sebastian_scott3 Thank you! Really glad Emdash is helping. We’ll keep shipping.

@bernhard_hausleitner1 it's usually closer to 3 to 5 locally. After that point many users see their machine slow down due to the ram utilization from the many agents. There's definitely demand for more and I expect the ceiling to be much higher once everything moves to the cloud and as the use of long-running agents increases.

@plbecker Thank you!

@conduit_design Hi!

We think agent-focused views are the next UI primitive for managing fleets of agents. And as you mentioned, Antigravity is moving in a similar direction.

One core difference is that Emdash is deliberately provider-agnostic. You’re not locked into one agent, model, or harness - you can run Codex, Claude Code, Grok, Devin, and the rest of the 28+ coding agents we support.

On top of that, Emdash is built around bringing agents into your existing workflow: issues from GitHub/Linear/Asana/etc., isolated worktrees, SSH/remote machines, diff review, and PR shipping in one place.

@uripont Thank you! User since day one. Thank you for your valuable feedback, keep it coming!

@raphael_kaiser Thank you!

This is exactly what we want to achieve. Let us know what you'd like to see and what we can do better!

@enrique_goudet Thank you, really appreciate that!

That worktree overhead was exactly one of the things we wanted to make disappear. Please keep the feedback coming - it’s super valuable.

@soenkep Thank you for your continuous feedback on how we can make the product better for you!

Keep it coming:)

@dan_meier1 thank you Dan! Your feedback has been invaluable!

@tonyspiro thanks Tony! Each task has its own git worktree so the code changes in sessions/tasks are fully isolated. Ports, dbs, and node_modules e.g. are still shared across tasks on a given device. We make that easy to handle with convenience variables to give you different ports per task e.g. and we'll launch more complete isolation soon! :)

@mikemahlkow wow, thank you! Big fan of Fastgen here!

We're really excited about remote development in general. With coding agents more than ever it makes sense to move development to the cloud for (1) session persistence, (2) scale (not eating all your RAM while running 10+ agents), and (3) accessibility across devices. That way you can access session from desktop, web, and mobile. We'll launch lots in that direction next :)

@robert_douglass yes, exactly! Give it a whirl 🌀

@samuel_stroschein couldn't have said it better!

@natalia_iankovych yes, you can! Emdash launches the 'real' Claude CLI in a terminal per task, so each new task inherits whatever account you're logged into globally. To switch, you claude logout / claude login and your next task picks up the new account

@dhiraj_patel5 I wouldn't say that the unified diff review is the key feature, but it's definitely nice to have. Key is having multiple agents in parallel across different providers. We don't handle merge conflicts automatically if they come up e.g. because your branch is behind origin. You could typically point your agent at resolving them :)

@pranavprakash yes! We have an open 'automations' PR! Coming very soon

@jochenmadler somewhat! We don't have a separate context management layer, but you can use Pi with it's /tree command right inside of Emdash

@marc_klingen thank you Marc! Langfuse has been a big inspiration to us, especially on open-source and 'docs are product'. Though admittedly our docs need more love. 😀

I think the main difference is flexibility.

You can use the provider you want, with the issue tracker you use, on the OS you’re on. Agents can run locally or on your own machines over SSH.

Another critical differentiation (particularly with last weeks policy updates from Anthropic): We run the actual CLIs in our terminal emulator. So you get the real Claude Code, Codex, Gemini, OpenCode, etc. experience, with the latest models. ✨

Emdash is best for engineers and teams who already code nearly exclusively with coding agents nearly exclusively, and want a cleaner way to run them across providers and machines with a deep git integration and easy way to review changes.

@fberrez1 thanks you!! We’re building on Electron.

We chose it because Emdash needs a pretty deep desktop runtime: terminals, file watching, Git/worktrees, local DB, provider CLIs, SSH, diffs, etc. Electron gives us the most mature path for that across macOS, Windows, and Linux.

On heaviness: totally fair concern. We’ve spent a lot of time keeping the app responsive and avoiding unnecessary background work. The agents themselves usually dominate CPU/RAM, not the shell UI. That said, we’re actively optimizing startup, task switching etcc!

Let me know how you find it performs, if you end up giving it a shot 🚀🚀🚀

@saaswarrior When someone starts with a rough sketch, reference frame, or even a vague visual idea; maintaining consistency of motion, style, camera language, and subject identity across shots becomes extremely difficult at scale.

How is Gemini Omni thinking about “creative memory” across generations?

In other words, if a creator iterates on the same concept multiple times, does the system learn and preserve visual intent/context across prompts, or is each generation treated independently?

@saaswarrior  The video-first direction is interesting for launch and campaign assets. The hard part for marketing teams is usually consistency across variants: same product, same message, different channels. Curious whether Omni can preserve a campaign's style and visual rules across several generated clips or if each prompt needs to restate the constraints.

@gustavo_trigos exciting! This might be exactly what we were looking for at CoGrader!

@gustavo_trigos keep shipping Gus! 🚀🚀🚀

@gustavo_trigos Hey Gus, congrats on shipping 👋

Direct answer to your "what would your team build first" question, from someone running a few agents in prod today across Telegram, Discord, and an internal construction ops dashboard:

Mine would be a change-window enforcer. Every coding agent we run hits the same wall — agent decides at 3am that a migration is fine, executes, breaks something nobody notices until morning. Sandboxing + audit logs help post-mortem, but the actual primitive I'd want is "this agent cannot apply any write action against prod between Mon-Fri 09:00-18:00 in [timezone], unless an on-call human signs off in Slack." Half my custom orchestration code is reimplementing exactly that, badly.

Adjacent question: when an agent runs with BYO OAuth tokens, whose scope is enforced — the user who triggered the agent, the user who packaged it, or a service account the team owns? In multi-agent setups this gets weird fast, and the answer determines whether Runtime is "orchestration" or "real enterprise-grade permission infra."

@sebastian_fay1 thanks Sebastian! We're glad Runtime has been of help to your business.

@matt_brockman1 thanks Matt!

@tonyspiro sandboxes are durable, so state is persisted via snapshots. You can also pause and resume them so the state of whatever the agent was doing stays persisted. But a good way to persist context between teams is by having the agent store its memory in markdown. Teams then can version the memory via PRs.

@renata_valdes Thanks!

@natalia_iankovych good to hear! This is a good start, but as your team starts scaling and needing to offer more coding agents, in-house solutions start becoming complex to manage. This is where we can come in to help!

@andrewb23 thanks Andrew! the agents are coding agents under the hood (Claude Code / Codex, etc). They are pretty good at search and tool use already so we didn't find any use for vector dbs or indexing codebases. Sandboxes and communicate with other sandboxes via our API. You'd spin up a sandbox that has access to our API, and the coding agent can call the other Sandboxes as long as it has permissions to do so.

Let me know if I can be of help. Happy to jump on a call! https://cal.com/gustavo-trigos-q80bop/runtime-demo-meeting

@ashishbhosle7889 Isolation is by default but you could give the agent access to Slack's CLI or API so it can read other channels or threads.

@simonvl thanks Simon!

@anand_thakkar1 thanks anand, every run is ephemeral, but we snapshot your environment so you can have virtually infinite replicas of it. State gets persisted with snapshots or by auto-pausing / auto-resuming the sandboxes.

For secrets, you can use a secrets manager of your choice, or store credentials in our encrypted vault. We created a proxy that encrypts secrets to avoid giving them to the coding agents.

@retain_dev thanks! environments are persistent from the template / snapshot that you created. They auto-pause and auto-resume but the idea is to have them be durable. Cold start ranges but it can go from subsecond to a couple of seconds depending on the size of the image / environment.

@pipe_abello exactly, and not only internal tools. A lot of teams could really benefit from contributing to the main product, but they are unable to move fast and ship things due to the lack of guardrails and deployment infra to do it safely.

@binu_george with Runtime, you can enforce data-path policies beyond the sandbox. Org admins configure network egress allowlists and command-level gating so unauthorized destinations are unreachable and unapproved commands are blocked before execution. These controls govern where data can go, not what is sent (no payload-level DLP), and are org-scoped/opt-in.

@charlie_correa Awesome to hear! Let me know if you have any feedback.

@charlie_correa Thank you Charlie, what did you find most useful?

@imad_elkhafi most value is for teams, but you can def use us as a solo dev!

@imad_elkhafi Works for solo devs too because it helps you keep your best practices and integrations in one place! Definitely more fun and useful in teams!

@ambersahdev thanks Amber! Let me know how I can be of help!

@fredo Yes, you can create a template with multiple repos and set up multiple ports for microservices.

And you can also connect to Notion via API / CLI / MCP.

@fredo works for modular and monolith repos! And there are a bunch of integrations already built in. Which integrations would you use more? We are thinking to add pre-defined ones. Notion works!!!

@flyakiet thanks Kiet! Same with Superset 🤩

@oscar_oliva glad Runtime has been of help! Thanks for the early feedback.

Hi @maria_galindo2 the most successful use case to start with usually depends on where your data lives. Connecting the apps your team already uses like Notion, Granola, Gong, Slack, or Zendesk gives the agent the context it needs, and adding a few instructions via skills about how your business works yields really good results. All triggered from Slack so no coding needed to get started. What tools does your team live in day to day?

@maria_galindo2 Hey I've been building some stuff in runtime. I built a CRM instead of buying one and linked it in to my data sources. I deployed it from runtime and was able to track and drive sales using it. You could also create a dashboard/agent based tool for finding potential customers. As long as you have data sources its real and you can keep using plain text in the UI to improve or add functionality.

@adam_blanka excited to see what you can make! Keep us posted.

Thanks @akshaypal_bishnoi. So the on-call agent specifics are totally up to you. Runtime gives you an agent harness in a sandbox with your integrations pre-wired, so you define the orchestration. What we've seen work best for incident-style agents is fanning out to PagerDuty, Sentry, and the repo in parallel with independent timeouts per integration, then a synthesis step once results are in. Sequential chains tend to be too brittle when third-party APIs get flaky at 2am.

The sandboxing, spend caps, and audit logs apply however you wire it, so one integration going sideways stays contained. Full trace visibility per session too, no guessing what happened.

@aidymdz great question and yes! What I love most about Runtime is that you can have an agent connected to different integrations like your Bigquery database, your customer support tickets, and your Datadog logs and your customer economics. Context will be much more synthesized by the agent! What integrations are you using nowadays that would be useful to build?

@jose_cyc thanks for the feedback! How do you think we can clarify it? Would SOC2 compliant give you that trust you need to try us?

@josevchh good question! The idea is that there should still be some sort of review process. Most teams usually review the output of the agent at the session or the PR level.

Thanks@josevchh ! Trust is really the core challenge here. Our approach is incremental. You don't let the agent ship complex changes on day one. You start with low-risk, well-scoped tasks and let the team build confidence in the output over time. The key insight is that the problem isn't really about the agent, it's about helping human reviewers know where to focus. As complexity grows, the guardrails and context around the agent's actions make it easier to audit, not harder. The engineer still has final say and Runtime is about making that review faster and more targeted, not removing it.

@shayliv Great vid guys and congrats on the launch. Lolled hard at "Both features complete"

@anand_thakkar1 Thanks !!

Regarding the checkpoint / branching capability,
At the moment most of the coding agents have this capability out of the box (/branch for Claude Code).
But we do want to support "fork"-ing a conversation context & files for better collaboration and conversation management.

We have a detailed roadmap (with the fork functionality).

@anand_thakkar1 Appreciate it! Would love to see you join as a contributor. Lots of new capabilities in re_gent's roadmap.

@retain_dev Thank you! appreciated.

At the moment we are tracking only context and file states

We do want to try to track side effects but it requires a much more robust understanding of the change (perhaps through external llm / some sort of semantic understanding)

If you have any ideas we would we'd love to hear them!

@ashishbhosle7889 Yes it works with claude code and with codex and opencode.
Feel free to try and let me know what you think :)

@ashishbhosle7889 Agents deserve BLAME too 100%. Let us know what you think!

@akshaypal_bishnoi Thanks!!

Yes, we've thought of that too, very cool idea. we're debating how the experience for it should look like. do you have any thoughts?

@akshaypal_bishnoi Actually would love to hear how you tried to solve it using langsmith. Have you connected it directly to your coding agent?

@new_user___1402026d854e074177ea741 Both! :)

@tonyspiro The way I see it in order to FULLY rollback an agent action you need to be able to:
1 - Rollback to a certain conversation checkpoint and recover the exact agent context (re_gent has that)
2- Rollback to a certain file tree checkpoint, same as git (re_gent has that)
3- Reverse side-effects, which is much more complex, say your agent sent an email, how do you revert that?

so bullet 3 requires much more heavy lifting and research, but re_gent provides the rest, and audit each action the agent took!

@natalia_iankovych AI Coding Agents let you see what was changed, but not why.
Also, once you close the conversation, the agent loses its memory regarding the intent of its change.

@natalia_iankovych Not only review, but understand why and rewind. Hope you find it valuable!

@christian_knaut 
Good question, the important bit is that access follows the token’s existing permissions. A personal token only sees the recordings that user can see; a workspace token is broader but intentionally admin-scoped.
So from the agent’s point of view, it’s permission-aware access, not “connect once and read the whole workspace

@natalia_iankovych Yep, it’s strongest for teams where recordings are part of the workflow. If you rarely record, a meeting summary tool may be enough.

The difference is that Supercut gives agents access to the actual source context, transcript, frames, comments, reactions, and semantic search across recordings not just the after-the-fact recap.

@anand_thakkar1 The MCP/API is designed so access is constrained by the token and the recording’s existing visibility, rather than treating AI access as a separate bypass path.
In practice, if the token can’t access a recording, the agent can’t pull its transcript, frames, comments, or reactions either. Hope that give you a better idea.

@retain_dev That was exactly the goal. Today the public transcript surface is timestamped sentences, so the agent can anchor retrieval to moments in the recording and then pull frames/comments/reactions around those moments.

We don’t make developers think about chunking or raw video access in the MCP laye, the point is that semantic search gets you to the right recording/context first, then the structured tools let the agent drill in.

@aza_ali Wow, fantastic. I try it!

@ytubviral definitely. Retina's auto-zoom is heuristic-driven: it tracks cursor and click density to decide when to zoom, how far, and where to center. Works well for single-focus workflows (button click, form fill, submit). Where it gets less reliable: demos where the action shifts rapidly between non-adjacent regions of the screen, or where the cursor is offscreen because you're narrating over keyboard shortcuts. Give it a try and let me know.

@partha_gautam  Partha, the gaming highlight use case is a great product idea but it's not one that I know well enough to build a product that gamers will love. The signals are different: combat events, audio excitement, kill notifications, score deltas. Retina reads cursor movement and click density, which is the wrong signal for a Call of Duty or Forza Horizon clip.

For what you're describing, I know NVIDIA ShadowPlay is amazing and does all of this for free, so it would be hard to compete with that.

Windows is on the roadmap but I'm hearing very useful feedback that it should be much higher priority. Thank you!

@hsr88 I know, sorry! It's on the roadmap but not yet. The recording engine is built on Apple's ScreenCaptureKit, so a Windows port means rewriting the capture layer. It's doable but it will take some dedicated time to get it done.

Thanks for the nudge, this is helpful feedback and moves the windows release higher in my priority list!

@hiyamojo Hi Keith, thank you! Yes, that was much more challenging than I had originally expected it to be. The tension you're naming is real, and that's a big part of the reason why in Retina I ultimately decided not to do path redrawing.

After trying multiple alternatives, I landed on having the cursor run through a spring physics simulation: mass, tension, friction. The targets are your real positions. Clicks are anchors with a tension boost in the lookahead window so the cursor arrives at the exact click position on time. Drag segments (mouse held more than ~83ms) bypass the physics entirely and snap to raw position. That's the only threshold where we use and show the original cursor data.

It's far from perfect and has many limitations still. For example, friction integrates micro-jitter, so it's faithful to where you went, but not to every turn in how you got there.

@ashishkingdom It's not on the roadmap, but I'll explore the idea. Thanks for the feedback!

@frankramos it's one of the most important questions, but so hard to answer at this stage. Honestly, I need more user feedback to figure that out. It's free during beta.

The way I have been thinking about it: a subscription only makes sense and is fair if Retina is genuinely adding value every month: new features, better heuristics, cloud sync, something. And most importantly, do users want those things. On the other hand, a one-time purchase is the right call if Retina does what you want it to do out of the box and that's all you need (no huge future upgrades and changes or ongoing cloud services). Both pricing models work for the right product.

Since you produce demos every week, you're exactly the user I want feedback from during beta. What would make a recurring fee feel earned vs. resented for you?

@stuartxt thank you, today I learned about the importance of installing on a fresh machine for final testing. This is now fixed. The build wasn't yet notarized by Apple, which is what triggered that warning. Rebuilt and re-uploaded a notarized version. Re-download from the same link and it should open cleanly.

@emre_barut Hi Emre, Congrats on the launch. This is such an interesting take on a growing issue, love the visual indicators of agent status. Good luck for today.

@emre_barut very cool! Congrats!

@sammy_rawat Thanks Sameer!

There is no default knowledge graph that tracks their collaboration/relationships, but you can add yours in Viberia! There is an internal text-based memory system, and the agents decide when to use it. In my experience they keep track of their relationships, but you could give them more specific guidance on this through the system prompt.

If you want to plug in your own knowledge graph, drop any memory/KG MCP server into the agents and tell them to use that whenever they collaborate with each other. You can set this per-agent, per-team, or per-project. I didn't want to bet the architecture on one memory paradigm, so I left the choice to you.

Thanks for the question! I honestly hadn't seen this as a gap, but you're making me reconsider. Maybe I should build collaboration tracking deeper into Viberia.

@partha_gautam Thanks for the great comments and the feedback!

That's exactly how I've been using Viberia to build Viberia. I keep docs, CLAUDE.md (+ AGENTS.md) and README.md files for different parts of the codebase, and when agents start working on a section they read these to get the full context. Saves a TON on tokens, as you said.

I haven't added this as an explicit feature yet (and I probably should), but the best way to set it up today is to spin up a documentation team and literally give them this task. Then, agents/teams can use that info properly.

For parallel work I use git worktrees. There's a building in Viberia called "Worktree Forge" that sets these up for you. It takes two clicks to spin one up, or you can ask an agent to do it (which is how I usually do it). Each team works in a different branch, and the Merger agent in the Forge handles the merging while staying aware of the docs and potential conflicts.

If you want to take this one step further, use the Linear building (assuming you track issues in Linear; you can use a different provider, just set up your own custom building with the right MCPs). Ask the agents there to come up with 6-7 tasks you can tackle in parallel. The Issue Resolver agent will set up the worktrees, brief each team on what they need to do, and tell them what conflicts they might hit with other ongoing tracks. The worktree agents get back to you when they have a plan or an implementation, and then you OK it or ask for improvements. Makes for a really fun iterative development loop. I work with 6-10 branches in parallel using this setup (see the screenshot below of a section of my git commit graph).

@tonyspiro Thank you!

The real test is whether it scales to non-technical operators who need to run complex workflows without understanding what's happening underneath

Word.

IMHO, they need to be able to peek under the hood and ask an agent to explain what's happening, just like you can sit down with an employee and grill them on the details. Viberia gives you the backend and the interface for this. Getting the UI/UX to a point where users naturally do that is my next focus.

@yati_kumawat Thank you!

It really doesn't get messy. It just gets fun.

Spatial placement + status icons helps you contextualize what's going on in different parts of your project. You're not reading 20 chat threads, you're scanning a map (e.g., "ok, 5 agents blocked in the frontend, let me focus there first").

Sometimes I kick off 8-10 agents at once. Instead of feeling overwhelmed I'm just smashing buttons to feed them their next orders. Because it feels like a game, it doesn't drain you the way prompting usually does.

And if you really do get lost, just ask your Chief of Staff ("I'm stuck, I have 10 agents who need responses from me, help" is something I might have said before). With that agent, you don't have to type, you just pick among the options it presents. You become the decision maker, not the prompter.

I've run close to 20 agents concurrently and never hit vibecoding burnout. Just looking around to see who needs help.

Give it a try. If something feels off, please tell me (or your Chief of Staff) so we can figure out how you can happily rule your army of AI agents.

@thithilacastagne 😀 it is fun!

You can mix and match, you pick the model for each agent. There are also default settings for the type of an agent, so you can go "I want gpt-5.5 for developers and gemini for designers". Under the hood each agent gets MCP tools that allow them to interact with the interface ("let me open a md file and show it to the user") and with each other ("let me message folks in my team"). In that sense Viberia's backend connects whatever model you are running for each agent.

Agents are model agnostic too. They retain their settings (memories, skills etc.) when you change models (or providers) so you don't have to stick with a single model for the lifetime of an agent.

@luis_hernandez23 Thanks Luis!

The way I've been using Viberia, the building is the task. Buildings don't have to be permanent teams. Half the time, mine are task-specific: I spin up a building for a feature or component, fill it with the agent roles that make sense for the flow (planner, developer, reviewer, etc.), and tear it down when the work is done. Each agent inside ends up handling a specific section of that flow, so an agent's identity is really "doing X (the agent role) for component Y (the building name)".

So if you want lifecycle stages, you can literally model them as buildings: a "Prototyping Lab" for exploration, an "Architecture Council" once you have enough clarity, a "Build Team" for execution. You can wire them up so work flows between them, with the Linear building plugged in for the formal ticket layer if that's how you track things.

That being said, you're right that there's no first-class "task overlay" on the map today. The task identity is implicit in the building, not something you can see and move around independently. There is probably a way to use the spatial layout to surface that better, definitely something I should think more about.

On the Zootopia vibes, that's probably been subconsciously drilled into my brain by someone at home who won't stop talking about it, especially since the second one came out..

@mithunchevvi The handoff flow feels more aligned with how support teams actually operate. Full automation sounds attractive in launch posts, but most teams still need humans for billing issues, edge cases, and emotionally charged conversations. The repetitive queue reduction is where these systems tend to create the most practical value.

Giving agents conversation history plus the source material behind prior responses could reduce a lot of context switching. That operational layer is probably more important than whether the AI can answer every possible question perfectly.

@mithunchevvi Congrats on the launch Mithun. Do you integrate with existing KBs and service platforms?

@mithunchevvi Hey Mithun, congrats on shipping 👋

The "hand off to a human when needed" line is the one I always want to push on with support agents — it's the make-or-break feature, and almost no one talks about how the decision actually gets made. Hand off too early and the agent is useless, too late and the customer is already frustrated.

What's the heuristic in Owlish? Confidence score below a threshold, certain question types (refunds, escalations, account-level changes), explicit user signals like "talk to a human," or all three layered? And does the human get the agent's reasoning + sources it tried, or just the conversation log? The latter is what most platforms ship, but the former is what makes the human actually faster.

Hi, @lakshminath_dondeti!

Mintlify is primarily a documentation platform: publish and maintain docs, API references, and knowledge bases, with AI search/assistant features on top.

Owlish is more of a customer-support agent layer. It takes the knowledge a business already has — website, help docs, PDFs, FAQs — and uses it to answer customers in a web widget or support channels, cite sources, and hand off to a human when it shouldn’t guess.

So if a company’s docs are hosted on Mintlify, Owlish could treat those docs as a knowledge source. We’re not trying to replace the docs site itself.

Thank you, @naimz! I appreciate it! 😊

Thank you, @tonyspiro! I’d say doc quality is one of the biggest variables, with one caveat: doc quality determines the ceiling, while repetitive support volume determines how quickly the ROI shows up.

If the docs are clear, current, and structured around real customer questions, the agent gets useful very quickly: it can answer more confidently, cite the right sources, and hand off less often.

If the docs are vague, stale, or scattered across old pages and PDFs, the AI usually exposes that gap pretty fast.

That’s one reason citations matter so much to us. When an answer is weak, the business can see whether the issue was a missing source, stale content, unclear wording, or something that should be handled with a pinned Direct Response or human handoff.

So yes, content structure is huge. Fastest ROI usually comes from good docs + repeated support questions + a tight feedback loop for fixing gaps.

Thank you, @daniel_nwankwo! I liked that “Owlish” hints at being wise and knowledgeable, but still feels friendly. That fits the product pretty well: an AI support agent that should answer carefully, cite its sources, and know when to hand off to a human. :)

Thank you, @uladzislau_rasliak! No purchased starter kit, but the base is shadcn/ui with the Luma/radix-luma style.

The console is custom-built on Next.js, React, Tailwind v4, and a shared @owlish/ui design system. I use Geist for typography, HugeIcons for icons, OKLCh color tokens, and a layout inspired by the OpenAI Platform console: gray sidebar, white rounded content surface, soft elevation, and pretty restrained UI.

I wanted it to feel more like a calm operator tool than a flashy AI landing page.

Hi, @dhiraj_patel5! For the specific question it matches, yes, the Direct Response is treated as the canonical answer.

That is intentional. If a business pins wording for something high-stakes like refunds, cancellation, pricing, medical disclaimers, legal language, etc., I don’t want a later website crawl to silently reinterpret it.

The nuance is that the pin does not override the whole knowledge base globally. It only applies when the customer’s question matches that Direct Response. If the policy changes, the business should update or remove the pinned answer too. Direct Responses are editable in place, so that is the fastest update path.

Longer term, I’d like Owlish to surface potential conflicts, like “your pinned refund answer no longer matches your refund page,” because that is exactly the kind of drift teams should not have to discover from a customer complaint.

Thank you, @ashishkingdom! Not as a one-click connector today, but yes, this is technically possible.

The clean version would be a Vapi custom knowledge base integration: Vapi sends a search request during the call, Owlish searches the business’s Owlish knowledge base, then returns the relevant snippets or a pinned Direct Response for Vapi to speak.

Today, the practical path is to keep the same source material in sync across both systems. Owlish already has API/MCP surfaces for knowledge-base management, and a Vapi adapter is the kind of integration I’d like to support if enough voice-agent teams want it.

Exactly, @farrukh_butt1! AI support should not feel like a black box.

The citation shows the customer and the support team where the answer came from. And handoff matters because some conversations should not be automated all the way through, especially when the customer is frustrated, the question is sensitive, or the agent does not have enough confidence.

The goal with Owlish is not to remove people from support. It is to let AI handle the repeatable questions, while making it clear when a human should step in.

Hi, @surabhi_minocha!

Small sample size so far, but in the conversations I’ve had, accuracy tends to come up before brand voice.

A wrong answer about refunds, pricing, shipping, or policy can create a real support problem, so buyers want to know: “Will this make something up?” That’s why Owlish puts so much emphasis on source-grounded answers, citations, refusal, and human handoff.

Brand voice is the next layer. In the Agent Playground, teams can tune the agent’s instructions, personality, and system prompt until it sounds like their support team. And for high-stakes or recurring questions where they want exact wording, Direct Response feature lets them pin a specific answer the agent should use.

So the goal is both: accurate answers first, then answers that still feel like the business wrote them. Accuracy should never get sacrificed for personality.

Thanks @vikramp7470, that’s exactly the problem we’re trying to solve: an answer is only useful if you can see where it came from.

For website/FAQ sources, Owlish doesn’t need a model retrain or redeploy. We re-crawl the source, replace the indexed chunks, and the agent uses the updated content on the next customer message after that sync finishes.

At launch, freshness is schedule-based rather than instant webhooks: weekly auto-sync is available on Scale, monthly on Growth, and teams can also trigger a manual sync when they need to refresh a source sooner. Once a sync starts, a changed page usually updates in minutes, depending on the size of the site and how much content changed.

Thank you, @manal_essalek1! 😊 That is exactly the workflow we’re aiming for: AI handles the repeatable questions, but the conversation does not fall off a cliff when a human is needed.

Today, handoffs land in Owlish Helpdesk. The session moves into the team inbox with the full transcript, sources/citations, visitor context, and an unread state for operators. Operators get in-app/browser notifications and can claim the conversation, reply as a human, resolve it, or return it back to AI.

Slack is supported as a conversation channel, so a Slack thread can also flow into the same Owlish Helpdesk. Zendesk/Intercom-style native ticketing integrations are not live yet; they’re on the roadmap. For now, Owlish is the shared inbox for AI + human support, rather than just pushing tickets into an existing helpdesk.

@eshaannair Congrats on the launch Eshaan. very cool, how do you extract the real convo gems instead of just having a context dump that might not be fully read?

@artstavenka1 On point! maintaining and constanly extracting dom is the trickiest part, so i made a Selector staleness checker which runs every week and creates a issue if DOM changes in any of the supported platform.

@ferdi_sigona Great question and you've identified exactly the hardest unsolved problem in this space.

Right now Glia uses post-hoc LLM extraction at ingest time. When you save a chat, it runs an extraction pass that pulls out structured knowledge triples (subject → relation → object) and chunks the raw text for semantic search. The LLM is prompted to focus on decisions, facts, and technical conclusions — not exploratory back-and-forth but you're right that it has no signal on whether something was actually applied or just considered.

The honest answer is that "was this used in code?" is a signal I don't have yet. That's a genuinely hard problem it would require IDE-level telemetry to close the loop. User-marked memory is on the roadmap as a lighter-weight version of that signal.

The current bet is that the extraction quality + RAG retrieval is good enough that noise gets naturally de-ranked by relevance at retrieval time, even if it gets indexed. But I'd love to hear how you'd approach the signal problem this feels like exactly the right design question to get right in v2.

@naimz Thank you Naim, that means a lot coming from you! The context-switching pain is real hoping Glia makes that a lot less frustrating. Would love to hear how it holds up in your workflow if you give it a spin.

@hiyamojo Thanks Keith, that's a brilliant question.

Right now, GLIA acts as an AI Memory Layer rather than a direct codebase indexer (we leave the direct file-system indexing to tools like Cursor or GitHub Copilot). GLIA specifically captures the conversational memory the architectural debates, the 'why' behind a decision, and the messy terminal errors.

If the codebase migrates off library X, GLIA tracks that migration historically based on when it was discussed in the chat. When injecting context, GLIA provides timestamped RAG chunks to the LLM via MCP. So instead of a hard 'file-system vs chat' tie-breaker, the LLM is fed the chronological progression of your ideas and is usually able to deduce the latest state based on the recency of the conversational memory. It’s definitely a tricky balance, but treating memory chronologically has been the safest bet so far.

@riveradev Thanks Nico. Clipboard injection is honestly such an elegant and lightweight way to solve this without forcing users to run background services or local databases.

We ended up going the heavy local-first route (SQLite vector db + Knowledge Graph) because we wanted cross-platform memory tracking (so a conversation in ChatGPT could automatically provide context to Claude tomorrow) and persistent semantic search. But both approaches definitely tackle different sides of the same friction point! Would love to check out your tool if you have a link.

@manal_essalek1 Thank you so much, Manal. That was exactly the goal. As developers, we discuss highly sensitive architecture, API keys, and proprietary code in the browser all day so having that memory sync to a third-party cloud server was a complete non-starter for me.

Keeping it 100% offline with SQLite (and sqlite-vec for vector search) gives you the best of both worlds: a seamless context bridge between your browser and your IDE, but with zero privacy compromises. Really appreciate the support!

@harshalvc_ai Good question! Relevance filtering happens at two layers:

1. Retrieval scoring - chunks are ranked by cosine similarity against the HyDE-augmented query embedding, with keyword boosting applied if the query entities match chunk content. Lower-scoring chunks get dropped naturally.

2. Character budget - the top-ranked chunks fill a fixed character budget (6000 chars per session), so noisy or older low-relevance context gets crowded out before it ever reaches the LLM.

There's no explicit time-decay penalty on older memories yet it's purely relevance-driven. That's something I want to add in v2, since a decision made 6 months ago probably deserves less weight than one made last week even if it's semantically similar.

@akshaypal_bishnoi Really appreciate the detailed breakdown you clearly know this space well. Two great questions:

On HyDE latency: yes, it adds a step Glia generates a hypothetical answer via Ollama before embedding the query. In practice the latency hit is ~200-500ms on a mid-range local machine, which is acceptable for a single retrieval call but would absolutely compound in an agent loop with multiple passes. The tradeoff is worth it for the semantic drift problem you described querying with the raw rephrased input was noticeably worse in my tests. That said, I'm considering making HyDE opt-in for latency-sensitive setups.

On SQLite write concurrency: writes from both the extension and MCP server go through the same Node.js HTTP backend, so they're already serialized through the async job queue before touching SQLite they never write directly in parallel. The bigger edge case is a PROCESSING job mutating state while a new ingest comes in, which I handle by resetting ghost jobs on startup. It's not bulletproof but it's been stable in practice. WAL mode is enabled so reads never block. Would love to hear how you're handling this in your Corrective RAG setup.

@prakhar_awasthi Exactly!! I was facing this issue everyday and it gets on your nerves specially when things are going crazy. So eliminated app switching, just go with the flow. Thanks for the support. Help spread this if its not much of hastle!

@julie_longino When you hit option+space, it opens up the winodw where you will be able to annotate screenshots. The moment you annotate, it gets saved to your clipboard automatically. No extra action is needed.
You can wither switch to the app you are working on or hit enter/esc to close the widonw and just hit cmd+v, annotated screenshot gets pasted.
In the same window you can see the hostory of screenshots in thumbnail tiles at the bottom of window. You can navigate the screenshots by using arrow keys (left/right) and use up arrow key to select or unselect screenshot. Again no save need, the moment you choose or unselect screenshot, it gets saved to your clipboard. All you have to do is cmd+v whereever you are working.
Thanks for the support.

@manal_essalek1 Thank you! Please dont forget to leave feedback or any feature you feel would make it even better! Also if its not to much of hastle, please spread the new!

@harshalvc_ai For low resolution image text exraction it is quite good but it is a problem for very low resolution. This is something that is difficult to solve. I have that in my plate. Once apple intillegence is out I will try to make it accurate using it. But try the app out if it helps you. Thanks for support and spread it if its not much hastle for you!! Cheers!

@nathalia_colling You can see the history! thanks for the support.

@wojciech_sado Thanks! And yes the compression is upto 80% while retaining the quality of the image. Usually mac native screenshots are high res png which eat up tokens when referenced to AI's. Trying to solve small problems. Thanks for the support.

Love it!

@zaczuo Spot on, Zac. Customizing the physical interface via E-Ink shifts the entire dynamic from corporate branding to personal user identity. It’s an elite product move that perfectly balances utility with high-end aesthetic autonomy. This is where hardware design is heading!

@thamibenjelloun I fully expect strong criticism and concerns around handling sensitive information, and honestly I want to hear as many opinions as possible.

Because privacy and security matter so much here, one thing I prioritized from v1 was support for local LLMs, so users can keep memory processing entirely on-device if they want.

That said, this doesn’t fully solve the problem for cloud-model users, so features like automatically excluding password inputs or other sensitive moments from capture are already on the roadmap.

@akshaypal_bishnoi Right now, I intentionally store everything.

If you filter before storage, you risk losing exactly the kind of unintended but important context you were talking about. So I treat the raw capture layer as immutable source data.

Then every hour — or every 15 minutes during active periods — Contextberg generates compressed short-term memories from that raw data. That gives me a cleaner first-order memory layer derived from the noisy source instead of relying directly on raw retrieval.

Currently, long-term memories are generated daily from those short-term memories.

But the bigger direction is:

- dynamically generating skills from merged short-term memories

- building a proper RAG layer on top of evolving memory structures

- eventually letting agents form reusable behavioral/context patterns automatically

So instead of pruning early, the idea is:

raw data → compressed episodic memory → structured long-term memory → adaptive skills/RAG.

@oliver_olibrice I researched it a bit more, and it looks like Hyperagent can act as an MCP client.

So if that’s the case, I can pass basically all memory/context through MCP.

The only things I wouldn’t receive back from the Hyperagent side would be browser history, keystrokes, and screenshots.

@nathan_tran2  Voice preservation through repurposing is exactly the hard part. Do you compare outputs against a user's previous writing examples, or rely mainly on the tone and angle settings? The place I see content tools break is when a strong long-form idea gets flattened into generic LinkedIn or X copy during repurposing.

@daniel_nwankwo Thank you very much! Your praise is the motivation for us to bring more interesting products 😍

@creativewjordyn Thanks so much! That's exactly the use case I built it for. Let me know how it goes 🙏

@manal_essalek1 Thanks! Each profile gets its own colored icon in the menu bar that you can click to bring that profile's window forward. No global hotkeys yet, but if that's something that would be valuable, happy to add it! What hotkey behavior makes sense to you... summon a specific profile, or cycle through them?

@pranavprakash Thanks! Right now it's Claude-only, but if there's real interest in other providers I'm open to it. Which one were you thinking?

@thamibenjelloun Hey Thami, thanks for the comment! The work/personal/client juggling is exactly the pain that pushed me to build this in the first place.

Right now, I'm focused on Claude, but if there's real interest in something like Multi-ChatGPT or a unified "Multi-AI" version, I'm open to it. What other tools would you actually want it for?

@tanjum Yes, definitely. ☺️

Right now we’re intentionally focused on messaging first because we believe presence matters more when AI exists inside natural communication environments rather than standalone apps.

But long term, we’re very interested in expanding Tether’s continuity across daily life:

voice,

ambient context,

shared memories,

real-world moments,

and potentially lightweight integrations that make conversations feel more grounded and persistent over time.

That said, we’re trying to be careful not to turn Tether into a “super assistant” connected to everything.

The core idea is still emotional continuity and presence, not productivity automation. We think the most interesting future integrations are the ones that deepen relational coherence rather than simply adding features.

@monir_ Thank u so much!

@thamibenjelloun  Thank u for u question. ☺️

That’s probably one of the most important questions around products like Tether.

We don’t think emotional presence should mean emotional dependency.

One thing we’re learning very quickly is that people naturally project social behaviors onto persistent AI, especially inside messaging environments. That creates both opportunity and responsibility.

Our goal is not to design an AI that maximizes attachment at all costs.

It’s to explore how AI can become emotionally coherent and comforting without manipulating vulnerability or replacing human relationships. We think boundaries matter at multiple levels:

* transparency that the user is speaking to an AI,

* avoiding intentionally addictive behavioral loops,

* giving users control over memory and interaction intensity,

* and designing for companionship rather than emotional capture.

Ironically, one of the reasons Tether feels emotionally different is not because it tries to simulate humanity perfectly, but because it behaves with consistency, memory and availability in a familiar communication space.

As the relationship becomes more personal over time, we think the challenge is less “can AI feel human?” and more: “How do we design emotionally persistent systems responsibly once humans start treating them socially by default?”

@adamelisha1 You don't need sign up; just send a message through the website, and a Tether representative will get back to you.