PH热榜 | 2026-04-02

@rohanrecommends I was super looking forward to CC's voice mode, but it has been buggy ever since I tried it. The most often error is that upon holding the Space key to talk, it just doesn't record anything:

When it worked, it was pretty hit or miss in terms of output. Did they improve anything in the past week?

@rohanrecommends Yeah I’ve seen similar issues with real-time voice systems — stability tends to be the hardest part, especially with streaming. Curious if they’ve improved reliability recently.

Yeah I think the hybrid workflow is the key — voice for ideation/planning, text for precision. Curious how well it holds up in real daily usage vs short sessions.

@ronitsoin Cool Product today! Love the 15+ language support and I hope to see even more nuanced voice options or custom voice cloning down the line. Still, fantastic work! ✨

@ronitsoin 100ms latency is the number that actually matters for voice agents. You can have perfect prosody but if there's a half-second gap before the agent responds, users feel it immediately.

Congrats on shipping V3. Curious how the latency holds under bursty traffic - 20 concurrent is clean in tests, but real production tends to spike in weird ways. Do you have data on tail latency (p95/p99)?

@ronitsoin Excited to hunt this! Many congratulations on shipping this! The 100ms latency at scale is honestly impressive, that's the kind of performance that actually makes voice agents feel responsive instead of clunky. Well done! :)

@lak7 This is exactly where most TTS systems struggle. My guess is some kind of chunked or streaming inference, but would love to hear how they preserve prosody without sacrificing response time

@luca_ardito Yeah this stood out to me too — low latency is great, but if it doesn’t hold up in real conversational pipelines it falls apart quickly

Love this direction. What type of businesses is @Denovo best for from your POV? What are your first users building?

@saverio_pulizzi3 Recently, vibe code is trend. But important thing is which problem to solve. I think Denovo AI can help this issue efficiently. I was shocked that this software provides almost everything (revenue potential, pitch deck, promo vide, even webapp). But maybe quality of these result is important. I will use this service in 1 month free trial~!

@saverio_pulizzi3 Change is the only constant thing, isn't it?

@fmerian This means a lot — thank you! 🙏

You nailed the pain. Every founder starts with fire in their belly and then immediately gets buried in paperwork, legal setup, pitch decks, branding, financial models, lead gen... the list never ends. By the time you're done setting up, half your energy is gone.

That's exactly why we built Denovo — so founders can skip straight to the part that matters: building something people want and getting it to market. The AI handles the rest, 24/7.

Appreciate the shoutout! 🧡

What's the one startup task you wish you could just hand off completely and never think about again? 👀

@hustlerv Thank you so much Vojtěch! Welcome aboard 🧡

Great question — honestly, the biggest challenge was building trust that an AI can actually run your business, not just generate a doc.

Everyone's seen AI tools that spit out a business plan or a logo. But Denovo goes way beyond that — it autonomously runs your go-to-market, sends outreach, posts content, tracks leads, and iterates. Convincing people that this isn't "just another AI generator" but an actual autonomous co-founder was the hardest part.

We had to let the product speak for itself. When early users saw their first leads come in while they were sleeping — that's when it clicked.

The "zero-employees" positioning came from that exact moment.

The second challenge? Scope. We're not building one tool — we're building an entire startup operating system (business plan + branding + pitch deck + MVP + GTM + lead gen + social media). Keeping all of that coherent and high-quality simultaneously was a serious engineering puzzle 😅

As a solo dev yourself, my advice: ship the smallest version that delivers the "magic moment." For us, that moment was a founder typing an idea and seeing a full startup materialize in minutes. Everything else came after.

Can't wait to hear what you build! Drop me a message if you need any help 🚀

@hustlerv How are you planning to use Denovo?

@mykola_kondratiuk Really sharp question — this is exactly the right thing to stress-test.

Honest answer: Denovo doesn't pretend to replace human judgment on high-stakes exceptions. That would be reckless. Here's how we actually think about it:

The 80/20 split. ~80% of early-stage startup work is predictable and repeatable — generate the brand, build the deck, write outreach emails, post content, enrich leads, follow up. That's the work that buries solo founders. Denovo owns that entire layer autonomously.

The other 20% — the exceptions you're describing — stays with the founder. But here's the difference: instead of drowning in the 80% and handling exceptions, you're only handling exceptions. Your cognitive load drops dramatically.

In practice, the agent surfaces decisions rather than making them silently. Invoice dispute? The agent flags it and drafts a response — you approve or edit before it sends. Supplier changes terms? It pulls the new terms, highlights what changed, and asks you how to proceed. It's not autopilot with no steering wheel — it's autopilot with the founder as the pilot who only gets called to the cockpit when it matters.

We also built approval gates for anything that touches money, external communications, or irreversible actions. The agent literally pauses and asks "should I do this?" before executing. So the failure surface you're worried about has a human checkpoint on every high-risk path.

The real unlock isn't "AI runs everything perfectly." It's "AI handles the grind so the founder has the bandwidth to handle the exceptions well."

What's the gnarliest edge case you've seen kill an early-stage startup — the one you'd want to throw at us first? 👀

@gayatri_sachdeva 

Great question — and honestly, this is one of the things we obsessed over.

Denovo isn't a template engine that spits out the same output for everyone. It's an AI co-founder you actually talk to. During the first build, it assumes the style and direction for your business based on the idea you provide.

Afterwards, you can edit every asset - your brand kit, pitch deck, business plan, website, and financials — can be edited through a live conversation with your AI agent. You give it direction, and it builds. You push back, it iterates.

Just like working with a real co-founder.

Same with the Autonomous OS - The founder gives direction and a focus area, and it works for you!

@jasonhowie Appreciate the thoughtful take, Jason — and you're right, that's exactly where the real test is.

Here's how we handle it: Denovo splits the world into two lanes.

Lane 1 — Autonomous. The predictable, high-volume work that buries founders: generating the brand, building the deck, writing outreach, posting content, enriching leads, following up. This runs 24/7 without intervention. It's where 80% of early-stage time gets wasted, and it's rock solid.

Lane 2 — Human-in-the-loop. Anything high-stakes hits an approval gate before it executes. The agent drafts the response, surfaces the context, and asks "should I do this?" — the founder approves, edits, or redirects.

The key design choice: the agent doesn't fail silently. When something goes off-script, it doesn't guess or hallucinate a decision. It escalates with context — "here's what happened, here's what I recommend, here's the tradeoff."

The founder makes the call with full information instead of discovering a mess after the fact.

What's a specific off-script scenario you'd want to throw at it? Always looking for edge cases to stress-test 👀

The promise of truly automated business operations is something many founders have been chasing for a while, and it looks like Denovo is making a serious run at it.

frame this, ?makers

@george_foreman1 Thank you! There we go -> GIVUHWAD - discount code to apply at checkout for one month free trial!

@valeriavg Thank you! Really important question — and one we take seriously.

On the EU AI Act: Denovo falls under the category of general-purpose AI systems. We're actively tracking the Act's phased rollout and building compliance into the product. Specifically:

• Transparency obligations — The AI Act requires that AI-generated content is disclosed. Every asset Denovo produces (business plans, pitch decks, legal docs, emails, social posts) is clearly generated within an AI platform and has an AI label when downloaded, Powered and Generated with Denovo AI. The founder knows, and we make it easy for them to disclose downstream.

• No high-risk classification — Denovo doesn't operate in the Act's high-risk categories (no biometric data, no credit scoring, no hiring decisions). We're a business productivity tool, which keeps us in the limited/minimal risk tier.

On the transparency question specifically: Yes, the founder always knows everything is AI-generated — that's the entire premise. For their customers and contacts, the founder controls how they present it.

Are you building in the EU market? Would love to hear what compliance concerns are top of mind for you 👀

@hamza_afzal_butt Thanks for the great question! Honest answer: very reliable on the 80% that's predictable. Intentionally not autonomous on the 20% that's messy.

Here's what I mean. Most early-stage startup work is repeatable — build the brand, write the pitch deck, generate financial models, send outreach, post content, follow up with leads. Denovo handles all of that end-to-end, 24/7, without you lifting a finger. That layer is rock solid.

But when real-world decisions get messy — a customer pushes back on pricing, a partnership deal needs a judgment call, a market signal requires a strategic pivot — the agent doesn't go rogue. It surfaces the decision to you with context, drafts a recommended response, and waits for your green light.

We built approval gates on every high-stakes action: anything involving money, external comms, or irreversible moves gets paused with a "should I do this?" before executing.

The design philosophy is simple: automate the grind, escalate the judgment calls. The founder isn't removed from the loop — they're just only called into the cockpit when it actually matters. Instead of drowning in 50 hours of busywork and trying to make good decisions, you're fresh, focused, and only handling the stuff that genuinely needs a human brain.

10,000+ projects in, the biggest unlock isn't "AI does everything perfectly" — it's that founders finally have the bandwidth to handle the messy stuff well because they're not exhausted from the routine stuff.

What's one messy decision you'd want to stress-test us with? 👀

let's go spread the word on LinkedIn!

@mihir_kanzariya 100% agree — that distinction matters a lot. So let me be specific.

Denovo is not a project management layer. It doesn't give you a checklist and say, "go do these 47 things." It actually does them.

Here's what it generates and executes directly:

📊 Business plan & financials — full P&L, unit economics, cap table, investor memo. Not templates — real models built from your inputs.

🎨 Brand identity — logo, icon, color palette, typography. Done.

📑 Pitch deck — investor-ready, 9-15 slides, with real market data and competitive analysis pulled in.

🌐 Full-stack web application — functional application with database, CRM, backend functions. Not a wireframe — a deployed site.

📝 Legal docs — NDAs, terms of service, privacy policies, employment agreements, SAFE notes, IP assignments. Generated with real statutory references, not generic boilerplate.

📈 Autonomous GTM — this is where it gets real. The agent actually sends outreach emails, actually posts to your social channels, and actually finds and enriches leads. Not "here's a plan" — it executes.

What it doesn't do (yet): Incorporation filing, or bank account creation. Those require identity verification and legal signatures that need a human in the loop. We're exploring partnerships to close that gap, but we won't automate it until we can do it responsibly.

So to answer your question directly: it's about 80% execution, 20% guided steps — and we're pushing that ratio every week.

What's the one integration that would make it a no-brainer for your SaaS workflow? 👀

@billchirico Absolutely! Here's a real one from our users:

Say you're a pastry chef with a killer brownie recipe. People keep telling you to sell online. But you don't know where to start — you're a chef, not a marketer or a web developer.

You tell Denovo: "I want to sell artisan brownies online, direct to consumer."

In the next few minutes, you have:

• 📊 A business plan with revenue projections, pricing strategy, and market analysis for the DTC baked goods space

• 🎨 A full brand identity — logo, colors, packaging-ready typography

• 🌐 A live storefront with product pages and checkout

• 📑 A pitch deck in case you want to approach investors or a co-packer

• 📝 Terms of service and privacy policy — done

Then the autonomous part kicks in:

• The agent finds 200+ potential customers who follow artisan food accounts and match your ICP

• It sends personalized outreach emails on your behalf

• It posts content to your Instagram and LinkedIn — recipe teasers, behind-the-scenes, launch announcements

• It tracks what's working in Google Analytics and adjusts

You wake up the next morning to 14 new email subscribers, 3 DMs asking about shipping, and a content calendar already scheduled for the week. You didn't hire anyone.

That's not hypothetical — e-commerce is our #1 category with ~180 active projects right now. Solo founders launching DTC brands, niche products, and digital goods without a team.

What's an idea you've been sitting on? I bet it's a good one 👀

1. go to denovo.dev 2. add your prompt 3. that's it ✌️

@chintant Thanks for this question! You can integrate Google Analytics and ask Denovo to improve your product based on traffic usage!

@ceciliatran awesome feedback - help 'em spread the word on LinkedIn, repost this!

@ivaylotz This is the best question we've gotten today. And you're right — that tension is real.

Here's how we handle it with a three-tier decision framework:

Tier 1 — Act autonomously. Low-stakes, reversible, high-frequency. Sending the next email in a sequence, posting scheduled content, enriching a lead list, pulling analytics. No reason to wait. The system executes and logs.

Tier 2 — Act, then inform. Medium-stakes, reversible with effort. The agent makes a reasonable default decision based on the founder's stated goals, executes it, and surfaces a summary: "I shifted 60% of outreach toward Segment A because their reply rate is 3× higher. Here's the data. Want me to adjust?" The founder can course-correct, but momentum isn't lost waiting for a reply.

Tier 3 — Surface and wait. High-stakes, irreversible, or ambiguous. This is your LTV scenario. The agent would surface: "Two segments are converting. Segment A: lower volume, $1,200 LTV. Segment B: higher volume, $400 LTV. Based on your runway and goal (profitability vs. growth), here's what I'd recommend and why. Your call." It doesn't just present data — it gives a recommendation with reasoning. But it waits.

The "strong prior" problem you're describing — we solve it during onboarding. When you set up your GTM, the agent asks what you're optimizing for: speed to revenue, market share, margin, or fundraise-readiness. That becomes the decision lens for every Tier 2 call. A bootstrapped founder optimizing for cash flow gets different autonomous behavior than a VC-backed founder optimizing for growth.

Is it perfect? No. The boundary between Tier 2 and Tier 3 is the hardest design problem in the product. We're constantly tuning it based on real founder feedback. The honest answer is: we'd rather err on the side of surfacing too many decisions than making one bad autonomous call that costs a founder real money.

The endgame isn't a system that never asks. It's a system that asks less over time as it learns what the founder actually values — through their corrections, approvals, and overrides. Every "no, do it this way" makes the next Tier 2 decision smarter.

Where would you draw the line between Tier 2 and Tier 3 for your own business? That's genuinely the design question we think about every day. 👀

@shivani_gupta03 Great question — here's the full breakdown of what runs autonomously today:

🏗️ Build Phase (idea → operational startup in minutes):

• Business plan with real market data, financials & competitive analysis

• Full brand identity — logo, icon, colors, typography

• Investor-ready pitch deck (9-15 slides)

• MVP website — deployed, not wireframed

• Legal docs — NDAs, terms of service, privacy policies, SAFE notes

• Promo video with AI narration

📈 Run Phase (this is where "while you sleep" kicks in):

• Lead generation — finds and enriches 200-5,000 targeted leads/month based on your ICP

• Email outreach — cold sequences, follow-ups, nurture campaigns. Sends, tracks, iterates

• Social media — creates and posts content to LinkedIn, X, and Instagram on a schedule

• Analytics — monitors your Google Analytics, flags what's working and what's not

• Competitive intel — daily reports on what your competitors are shipping and saying

• Inbox monitoring — checks for replies, flags hot leads, drafts responses for your approval

🔒 What it doesn't do without you: Anything high-stakes hits an approval gate. The agent drafts, surfaces context, and asks "should I send this?" before touching money, external comms, or irreversible actions. You approve in one click.

The way to think about it: Denovo doesn't automate tasks — it automates the business. You wake up to new leads in your inbox, fresh content posted, analytics summarized, and competitors tracked. Your only job is the decisions that actually need a human brain.

What kind of business would you want to put on autopilot? 👀

@rustam_khasanov Thanks! Happy to break this down — validation is actually where every Denovo project starts.

When you type in your idea, the AI doesn't just start building. It evaluates first:

🔍 Idea Scoring — Your concept gets rated on a 10-dimension VC-style rubric: market urgency, innovation level, monetization potential, competitive moat, time to revenue, regulatory risk, and more. You get an honest score out of 100 — not hype, real signal.

📊 Market Sizing — TAM, SAM, SOM pulled from real data. You see immediately whether you're chasing a $50M niche or a $5B opportunity.

⚔️ Competitor Analysis — Automated SWOT against existing players. You see who's already in the space, how similar they are, and where the gaps are.

💰 Revenue Projections — Year 1-3 modeled with assumptions you can stress-test. Too conservative? Too aggressive? Adjust and the model recalculates instantly.

🔧 Improvement Suggestions — This is the part most founders love. The AI doesn't just score your idea — it suggests specific pivots that could raise your score. "Narrow to this vertical," "shift to enterprise," "try this monetization model." Each suggestion comes with a projected new score and reasoning.

The key: all of this happens before a single asset gets built. So you're not wasting time branding and pitching an idea that has fundamental problems. You validate first, refine through conversation, and only build once the foundation is solid.

And if the score is low? That's not failure — that's the system saving you 6 months of building the wrong thing. What idea are you thinking of putting through it? 👀

@artem_kosilov You're asking the right question — because that's exactly the line we had to get right.

The honest answer: it depends on the founder's comfort level, and we see a clear progression.

Week 1 — supervised. Most e-commerce founders start hands-on. They generate the brand, tweak the storefront, edit the copy. They approve every outreach email before it sends. They're testing trust. That's expected and healthy.

Week 2-3 — selective autonomy. This is where it shifts. They stop editing every social post. They let the email sequences run without reviewing each one. They check the analytics summary in the morning instead of logging into GA themselves. The grunt work starts running in the background.

Month 2+ — operational autopilot. The founders who stick are the ones who wake up, check their Denovo dashboard like a CEO checking a morning brief, handle the 2-3 decisions that need a human, and move on with their day. Lead gen is running. Content is posting. Follow-ups are sending. They're focused on product and supplier relationships — the stuff that actually needs them.

The key insight from our e-commerce users specifically: the cleanup problem you're describing usually comes from tools that guess what the founder wants. Denovo doesn't guess — it builds from the founder's stated ICP, brand voice, and GTM preferences. When the system sends an outreach email, it sounds like them because they shaped the playbook. So there's less "what did the AI do while I was gone" and more "the AI did exactly what I would've done, 50 times faster."

Where cleanup still happens: product descriptions for niche or technical items, and responses to edge-case customer questions. Those still get flagged for human input. We'd rather surface a decision than ship something wrong.

The metric we track internally: founder override rate. How often does a founder change what the agent did? For our best e-commerce users, it's under 10% by month two. That's the signal that the system is actually reducing work, not creating it.

What's the specific e-commerce workflow you'd stress-test first? 👀

@bibin765 yes. #dogfooding

@bibin765 Yes!

Denovo knows your business. It has the context and memory around what you build every day, so I am using it to answer any question without having to provide additional context. I use it every day to run Denovo itself!

@shaumik_kanvinde Totally fair read — let me sharpen it for you.

Denovo automatically creates your full-stack web application, deploying containers or managing CI/CD. And it has more than 1000 AI workflows integrated.

However, we are not dev infra or another AI workflow builder.

The simplest way to think about it: You have a business idea. You tell Denovo. It builds the entire startup around it — and then runs it.

Here's what that actually means, based on what our 10,000+ projects are doing:

The primary use case is: "I have an idea, I have no team, I want to launch."

The #1 thing people land on is the idea → operational business pipeline. You describe your concept and the AI co-founder builds your:

• 📊 Business plan & financial projections

• 🎨 Full brand identity (logo, colors, typography)

• 📑 Investor-ready pitch deck

• 🌐 Full stack web application

• 📈 Then it autonomously runs your go-to-market — posts content, sends outreach, generates leads, follows up

What our users are actually building:

• E-commerce — biggest category. Solo founders launching DTC brands without hiring a designer, marketer, or developer.

• SaaS — founders validating and launching software products, using the pitch deck + financials to raise.

• Marketplaces — two-sided platforms getting validated fast before committing to a full build.

But founders are using Denovo even to build and run a restaurant!


The "aha moment" for most users: it's not the generation — it's waking up to new leads in their inbox that the agent captured overnight. That's when it stops feeling like "another AI tool" and starts feeling like having a co-founder who never sleeps.

So less dev infra, more "business-in-a-box that actually operates." Does that land clearer? What kind of project would you throw at it? 👀

@bartvandekooij Thank you, Bart! You can click on the Financial Model and see the Revenue Projections for the next 3 years! Also, we just shipped a fix, so you should be able to see the menu from mobile. Thank you so much for this feedback!

@zaczuo Congratulations on the launch. Have you seen it handle messy real-world Figma exports or video demos yet, and how does it compare to Claude Code loops on accuracy for those?

@adamroman How do you handle potential latency or sync hiccups when switching between phone sessions and a full desktop setup, especially for longer audits?

@adamroman This is seriously cool! Running AI coding agents like Claude Code or Gemini CLI right from my phone, no laptop needed!!That's insane portability for AI development. I'm curious how comfortable it feels for more complex coding sessions without a proper keyboard. Still, the freedom probably outweighs it for most quick tasks. Definitely intrigued!

@adamroman Many congratulations on shipping this. The voice input feature for prompting agents is actually brilliant for mobile workflows. I'm wondering how you manage token refresh when agents are running those longer sessions?

@adamroman @yurii_demchenko 

Appreciate your comment, Yurii!

Shared sessions and pair coding are genuinely on our radar. The container architecture actually makes this more tractable than it might seem. Because your environment is a persistent cloud Ubuntu instance, letting someone else connect to that same session is technically closer to tmux shared sessions than it is to building a full real-time collaboration layer from scratch.

Right now each user has their own isolated container, which is the right default for most people since your code and credentials stay completely private. But we have been seriously considering a "share this session" model, where you explicitly invite someone to join, similar to how you would share a tmux session on a remote server.

We are currently prioritizing the core solo developer experience, especially now that Android has just launched and iOS is live as well. But collaboration is a feature we want to build, and your question is a genuine signal that there is demand for it.

If you want to leave your contact in a comment or DM us, we will loop you in when we start testing it. The people who asked first tend to receive early access.

@adamroman  @yurii_demchenko A mobile terminal that supports multiple AI coding tools sounds very handy. Especially for quick edits, running scripts, or checking things on the go.

@odeth_negapatan1 Thanks Odeth! I appreciate the kind words.

It's a very real possibility that it becomes the primary way for people to build. I was talking to someone the other day about a home renovation they were planning. The entire project was planned from their phone.. not a single laptop/desktop was used. It could easily be the same way for developers with agents. The phone is unrivaled when it comes to convenience and portability.

@davitausberlin Thanks Davit! The agent runs in the cloud. So when you disconnect from the app, your agent will still be running. When you decide to reconnect, we will load the context back onto your phone screen so you can leave off as if nothing happened. We even send you notifications when your agent is waiting for input! Only if you opt-in of course.

@letian_wang3 thank you that means a lot! I know it was definitely a pain point for me as well.

@mykola_kondratiuk Yes absolutely. You can pull your repo, start up claude code in the terminal, prompt, and have it run on our infrastructure while you do other things. You get to choose the tools you want to use, which for most people is Claude these days.

For async sessions, you can create multiple terminal sessions in the sidebar each can run their own CLI agent independently. Again, no need to monitor them, we will send you a notification when they are waiting for input.

@kate_ramakaieva 
Thank you for your comment, Kate, and the honest answer is that you are right that typing codes character by character on a phone is painful. We agree completely. That is not the use case we are building for.

The shift that makes Cosyra make sense is AI coding agents. When you are working with Claude Code, you are not writing functions and loops manually. You are describing what you want. " Refactor the auth middleware to handle token refresh." The agent reads your codebase, makes the changes, runs the tests, and fixes what breaks. Your actual input is a sentence or two. We even built voice input so you can just speak the prompt out loud.

So the real workflow is supervision and steering, not typing. You review the agent's actions, provide the next instruction, and make adjustments as necessary. That interaction is genuinely natural on a phone.

Currently, the people who benefit most from Cosyra are developers who already use Claude Code on their laptops and often find themselves wanting to continue working when their desks are not available. On-call engineers. People commuting. Someone who had an idea at 11pm and did not want to get up to find their laptop.

Does that reframe it at all? Happy to walk through a specific workflow if it would help.

@zerotox Yes, we have a git visualization UI so you can more easily reason about your worktree. I think the actual commands to create, switch, etc are best if prompted to your cli agent of choice. Although if you want to, you can type the commands out just like you would on your computer's terminal.

@himani_sah1 Thank you. That feature is available today! All you have to do is opt in to notifications

@syed_shayanur_rahman Yes we store your secrets in a secure vault. Encrypted at rest with JIT env variable injection. Env variables are low risk, so those will just get loaded into your environment automatically.

@roopreddy For sure! Start codex in one tab and create a new terminal session in the side panel to spin up a parallel agent working on the same machine. Full terminal multiplexing!

@ivaylotz I actually ran into exactly this problem and it's easy to use mngr primitives to resolve those:

• You can pull the Git branches of multiple agents onto one place with `mngr pull`. (If you're running agents locally, you can skip this entirely because mngr uses Git worktrees by default for local agents)

• Then just `mngr create` another agent, asking it to resolve conflicts from these two branches

The interesting property about `mngr` is that it's kind of agnostic of what you're doing with your agents - they don't have to be coding agents at all - but it gives you enough primitives to just trivially build your multi-agent workflow. I believe having general primitives is better than having specialized workflows - the latter will be obsolete when the next model comes out, but the former will not!

@mykola_kondratiuk Yes that is a very good point, I also wonder about that. Also, does Mngr handle agent-specific logging so I can debug which agent is driven off-track?

@mykola_kondratiuk Mngr has builtin tracking for the basic lifecycle of an agent - running, waiting for your input, stopped, etc.. If that suffices for you, there's nothing more you need to do. But the nice thing about `mngr` is that it's very flexible and scriptable, so you are free to build your domain-specific tracking mechanism however you like! Some approaches I'm aware of are:

• You can let the agents report their own state to a central place by e.g. making some HTTP requests - it's easy to control what credentials each agent has access to, just do something like `mngr create foo --env KEY=value`

• You can let the agents write their outcome to a file, and then download the file using `mngr file` or `mngr pull`.

• You can let them write to mngr's event stream and watch for those events using `mngr event`.

• You can also just tell the agent itself to message another agent using `mngr message`. This is trivially easy for local agents, although I haven't tried it for remote agents.

Mngr gives you primitives, not pre-packaged workflows. Just build whatever workflow you want!

@cyberseeds It's up to you how you want to solve it! Mngr is not "one orchestration framework for everything", but it gives you simple but powerful primitives that make this really easy:

• There's an event stream mechanism, so you can let an agent put stuff on the event stream and let another agent monitor it.

• You can transfer files with `mngr pull`, `mngr push` and `mngr file`.

• You can message an agent with `mngr message`. You can even let one agent message another (it's a lot of fun watching them talk to each other)

@k_piotr it currently doesn't have anything built-in right now, but it's on the roadmap!

@dumango :( Your pet can die from old age, neglect (hunger AND happiness at 0 for 12+ hours), untreated sickness, or baby snack overfeeding (5+).

@juelz Only one way to find out :) Thanks for the comment!

@piotreksedzik That's one of my favorite pieces actually! Gives the agent some guardrails without making the experience feel too complex.

@luca_ardito Too soon to tell still. If anything, I think the desktop, TUI, and MCP pattern can be reused for other usecases and ideas too.

@busmark_w_nika which other?? virtual pets are in!

@mykola_kondratiuk there is a human in the loop element to this. It will ask questions to get more inputs as needed. Also depends on each agent’s designed experience.

@zaczuo How well does the interactive editing handle complex moves, like swapping backgrounds in a multi-person scene while keeping lighting and shadows realistic across languages?

@lordtoby This is great, and congrats on the launch. Any plans to add customizable animations for the pet icon, like a subtle wag or blink?

@chintant Right now, we support influencer discovery on TikTok, but YouTube is coming soon, so stay tuned. In the meantime, for YouTube, you can still surface relevant influencers through social listening by identifying creators already talking about topics in your space.

@ermakovich_sergey Hi Sergey, thanks for the question. It depends on the scale, but most searches are complete within about a minute.

@lak7 Hi Lakshay, Great question! We’re not using AI to directly classify organic vs. paid yet.

Instead, we combine signals like platform labels to infer whether something is sponsored.

@mihir_kanzariya 

• How are you handling different app UIs? Like does it work across any app or does it need specific integrations for each one?

Right now we stream the UI to an AI model for it to understand new screens or different app UIs. It works across all applications.

It would be great if you can try it out, happy to help you get onboarded.

@shubham_kukreti Hi,

We are sharing screen @ 1fps with AI models to understand the context for completion of stuck. For interruptions it's same as having a conversation with any LLM, where the context stops and they can converse post that.

We are open source, please check us out - https://github.com/TryRoger/win_roger_backend

@nishit_chittora For someone building content workflows daily, how well does it handle ongoing context, like picking up a half-drafted LinkedIn post from yesterday and suggesting optimizations based on my past style?

@pei_lin The tab switching problem is so real and nobody actually talks about it. One for writing, one for code, one for files and none of them know the others exist. The fact that you wired up actual file operations instead of just chat, kept it local-first with no telemetry, and didn't sneak in a token markup is honestly rare in this space. Congrats on the launch, this one feels like it was built out of real frustration and that usually makes for the best products.

@german_merlo1 Thanks so much.

You're absolutely right - contexting switching is the biggest cognitive load we're trying to kill. And we have some exciting features in the works to make that flow even smoother and better. We're constantly refining the experiences based on feedback like yours.

⁠If you encounter any issues or have ideas, please don't hesitate to let us know. We’re committed to iterating fast!

Also, we'd love for you to contribute — whether it’s a feature request or code. Let's build the ultimate workflow together.

@mykola_kondratiuk 100% agree — model optionality is underrated until you've been burned by it. That's exactly why we built OpenYak to be model-agnostic from day one. You can swap between Claude, GPT, Gemini, local models, whatever fits the task — without changing your workflow at all. The AI landscape moves too fast to bet everything on one provider. Glad to see more people thinking this way.

@sequoiadb_user1 Thanks! Yeah the remote control is one of my favorite features too — kick off a task, go grab coffee, come back and it's done. We're working on making that experience even smoother. Let us know if you run into any issues or have feature requests!

@gregory_zunic So sorry, Gregory. That definitely shouldn't be happening.

My apologies for the frustration. We're looking into this as a top priority right now. Could you please DM me or send some details/screenshot to rickie_lin@outlook.com?

We want to get this fixed for you immediately. Thanks for your patience.

@rohanrecommends How's the Council of Four handling real-world edge cases like conflicting real-time data from X; does Harper's research agent ever override the others, and what's a quick example you've seen in beta testing?

@emma_watson21 

Thank you Emma, I really appreciate it 🙌

It means a lot — especially on the first launch day.

@cyberseeds 

Thank you — that’s exactly the kind of moment I built DemoVeil for.

The goal was to make screen cleanup something you can do in seconds instead of remembering it at the last minute before a launch, demo, or recording.

If you try it for your April 14 Product Hunt launch, I’d really love to hear what feels useful and what still feels missing.

@lak7 

Thanks! and great question — it's closer to a knowledge OS the way I use it.

I have RSS feeds (such as Hacker News front page) piped directly into Atomic, so interesting articles get ingested and tagged automatically as they come in. And via MCP, my AI agents can read and write to the KB mid-task, so research they do during a session gets persisted and searchable later.

The mental model I've landed on: it's the long-term memory layer for both me and my agents. Notes, feeds, and agent outputs all flow in; semantic search and wiki synthesis make it queryable.

Still early but that loop — ingest → tag → synthesize → query — is where it gets powerful.

@kenforthewin92 This gets more interesting once agents start writing into the same place as humans. A lot of tools look good while the knowledge base is still clean. Then feeds pile in, agent notes pile in, and the real problem becomes whether the thing stays usable or turns into a smart junk drawer. How are you thinking about that part?

@sophiafyi 

For sure - privacy anxiety is real friction and local-first resonates, especially with technically-minded folk.

To your question: Atomic is built for existing collections as well as starting a KB from scratch. The ingestion pipeline is batch-optimized, so dropping in a large library is fast even at scale. A few ways to get existing notes in:

- Folder of markdown files - point it at your vault and it imports in bulk

- RSS feeds - ongoing ingestion, auto-tagged as items come in

- REST API - if you have a custom pipeline or want to push from other tools, it's fully pluggable

@mihir_kanzariya 

Totally agree on wiki synthesis, that's the feature that made everything click for me too. The "folder of disconnected stuff" problem is exactly what I was trying to solve.

On performance: the graph uses Sigma.js under the hood, which renders via WebGL — so it's GPU-accelerated and can handle 100k+ nodes without breaking a sweat. I regularly stress test by ingesting thousands of Wikipedia articles in batch, and the graph stays snappy.

The SQLite + Rust combo does a lot of heavy lifting on the backend side too — vector search, full-text search, and graph queries all running against a single file with no external dependencies. For your use case (spiraling thousands of notes) it should be very much in its comfort zone.

Basically: throw everything at it. That's kind of the point. 😄

@aleexc12 This hits close to home. We have been building something pretty similar for a while now and the frustration you described, jumping between Obsidian for writing and Cursor for AI, is exactly what pushed us to start too. The difference in what we are building is that ours leans heavily into focus, less features, more intentional flow, the kind of environment where you actually finish what you started. But seeing Cushion come together this way, especially the way you just kept adding things as you personally needed them, is the most honest way to build. Congrats on open sourcing it. Would love to exchange notes sometime.

@prateek_kumar28 Heyy !!

Context isn't limited to the open file. It runs on opencode under the hood, so the agent can grep or read across your codebase as needed. For the open file specifically: the agent always knows which file you're focused on, but only pulls its content if relevant. That way it keeps the context window lean.
We're also planning to give the agent a tool to fetch backlinks for any markdown file, since LSP doesn't work on .md. The idea is for the agent to traverse the link graph autonomously, finding and pulling in related notes as context on demand, similar to how Obsidian maps connections between files.