PH热榜 | 2026-05-27

@dima__eremin This feels like the missing layer between great conversation and useful follow-up. Love how Bluedot makes real-world moments instantly actionable instead of lost in memory or messy notes.

@dima__eremin One question: when you tap to record on Apple Watch, how does Bluedot handle background noise and privacy in crowded spaces, and what's the typical latency from ending the recording to seeing it synced in CRM/Notion with Claude-ready context?

@sivaram_grandhe we support most of asian languages

@anthony_adams_ yeah, it's a good question.

@anthony_adams_ Yeah, 100%. Consent is a big part of this. We’re keeping it user controlled, with clear recording states and the ability to decide what gets saved or shared after the conversation.

@lakshminath_dondeti that's why apple watch integration is the best for it. It's minor impact on the battery.

@emma_watson21 thank you so much!

@navin_kumar_singh exactly!

@nikita_jain18 Totally! that’s exactly why we added RAG and MCP, so people can find the right context without digging through everything manually.

@rocky_bhai39 we do exactly this way!

@frances_diazon long -session memory is what we optimise for!

@nithin_raju1 Thank you! Yeah, wearables feel perfect for capturing things away from your desk.

@elnur_atakishiyev yes, easily!

@jim_jeffers Yeah, exactly. The hard part is not recording more, it’s helping people understand what’s actually worth keeping and why.

@mythic_dd Thank you! Honestly, the hardest part is reliability, making sure the recording works when you need it, and that people don’t forget to stop it after the conversation.

@james_dunnigan Exactly! No one wants to update CRM after a long day.

@michele_di_brigida_bajara we optimise well for the battery usage, so you should't worry. Your exact drain would depend on the mode of the phone.

@tonyzhangcy  @xin_chen17  @michael_t_chang  @hunter_powabase 

Congrats on the launch — the FinanceBench number is interesting, would love to know more about the retrieval setup behind it.

One thing I keep running into building on top of LLMs: the infra layer you’re describing solves the "get data in front of the model" problem really well, but structured multi-step pipelines still tend to fall apart at the schema validation layer — the model returns something that looks right but breaks the contract with the next stage.

How does Powabase handle that? Is there a way to define pre/post conditions on workflow steps, or is correctness mostly enforced downstream by the application?

@tonyzhangcy  @xin_chen17  @michael_t_chang  @hunter_powabase The "infra glue was the slowest, most expensive part" framing is the honest reason this category exists, nobody's AI app died from a bad model, they died stitching six tools together. Since token-efficiency is the thread you keep pulling on, that's where I'd push: FinanceBench is clean and bounded, but a client's real knowledge base is messy, duplicated, and growing weekly. The hard part isn't retrieval quality on a benchmark, it's what the cost curve does at 10x the documents. Naive RAG either bloats context and quietly kills the token-efficiency claim, or over-prunes and misses the right chunk. So: as a project's corpus scales, does retrieval cost stay flat, and does precision hold on dirty data? That curve is the difference between a demo scoring 98% and a backend teams trust in production, and it's the one number that'd make "token-efficient by design" provable rather than aspirational.

@tonyzhangcy  @xin_chen17  @michael_t_chang  @hunter_powabase Congrats on the launch, team. Pointing out that Powabase extends from Supabase is a brilliant positioning anchor because it immediately secures baseline developer trust.

My only tactical critique is that your main headline, "Postgres, RAG, and agents. One backend." reads a bit too much like a feature checklist. With massive live traffic hitting you today, anchoring your hero copy directly against the maintenance nightmare of gluing a standard database to isolated vector indexes and custom agent runtimes would drive a much higher activation rate.

Rooting for you guys today.

@tina_chhabra Thanks for confirming our experience! A lot of the custom gluing depends on the actual project specifications. If the business logic is complex and many third party integrations are needed, you might still need a good amount of glue code regardless of what backend framework you use.

Powabase's ReAct multi-agent orchestrations allow you to connect agents with custom tools and MCP servers though, so if you're building an AI app, hooking your agents up with tools is a lot easier than building everything from first principles.

Happy to hear more about what you might be working on!

@tina_chhabra thanks for the support! We've already been dogfooding Powabase for other internal projects and it has definitely sped up development

@studentzuo Great observation! Powabase has a full suite of Observability tools built in. It gives you a variety of information ranging from token usage breakdown by model to failed RAG indexings and LLM queries. Since it is a BaaS, every configuration and action is tracked in a Postgres DB table for your records. You can also build custom dashboards on top of auto-collected data to analyze it as you see fit.

Like Supabase, our Postgres db has row level permissions. For certain out-of-the-box agent tools (e.g., database_write), the user must grant explicit permission to the agent on a per-table basis before it can manipulate it. If you want human-in-the-loop workflows like Claude code, you can set up hooks in your agent orchestration so it waits for user's confirmation before continuing.

Would love to hear more about what projects you have in mind!

@studentzuo Thanks for the support!

@campritchard thanks for the support!

@campritchard Really appreciate the encouragement!

@1lastshot thanks for your support! Powabase essentially evolved out of GPT-trainer. We learned a lot of lessons when we tried using it as an AI agent platform for other apps we were building (e.g. AI-powered document processing) but ran into pain points with GPT-trianer's focus on chatbots. We're hoping that Powabase, "backend we wish we had", will help others build AI-native apps too, and we'd love to hear more about what apps you may be trying to build!

@1lastshot Following up from @michael_t_chang , Powabase was inspired by the fact that many users of GPT-trainer requested direct API access to the backend. Instead of stripping out the backend from an end-user SaaS, we thought it was best to properly address the root demand and build a dedicated AI-native BaaS instead.

Thank you for your positive feedback and encouragement!!

@hunter_powabase  @adam_foresman  Thank you so much! Ryden Solutions has been one of the most rewarding projects we've worked on — what your team is building is exactly the kind of regulated-industry use case that pushed us to take Powabase's primitives seriously. We have learned a ton from working with you. Excited to keep building together!

@adam_foresman Thanks a lot Adam! Your positive feedback made my day!

Powabase is perfect for use cases like Ryden. It's ideal for anything document-heavy and privacy-sensitive.

Looking forward to our continued partnership!

@daniel_henry4 Glad you noticed!

Two ways we impact token efficiency:

1. Building every project from first principles and debugging glue code+infra setup via Codex, Claude Code, or OpenCode can be VERY wasteful in terms of tokens. You can save significantly by leveraging well-crafted abstractions during development. This is where Powabase shines!

2. Traditional RAG or agent orchestrations done via Supabase + LangChain does not incorporate intelligent "agentic" retrieval. To get the right answer on the first try, you generally have to compensate with larger reserved context. This is wasteful in terms of tokens. Powabase's native multi-agent orchestration uses a similar architecture as Claude Code. When set up correctly, your agents can be much smarter about where to look for the most relevant pieces of context instead of stuffing everything in one go based on cosine similarity. You can also control reasoning levels directly if you want fewer "hidden reasoning tokens" at inference time.

@daniel_henry4 Adding to Hunter's point — the unglamorous lever here is observability. Most "hidden" orchestration cost is hidden because the existing stack doesn't show it. We surface token usage broken down by model, agent, and source (agent run / orchestration / workflow), plus per-run trace with reasoning steps, tool calls, and the RAG context that got retrieved for each step. Once you can see it, the optimization is usually obvious — you stop arguing about prompt length when the real waste was a poorly-targeted retrieval call.

@fberrez1 Yep, we are using pgvector under the hood. For multimodal content, we first convert the data into a text-based description using a VLM like Mistral OCR, then index the text. At retrieval time, the database searches the indexed text but pulls the original multimodal content (e.g., base 64 image) and feeds it as context to the LLM.

Agent layer does support tool-calling across multiple data sources. Out-of-the-box tools interface mostly with built-in capabilities of the BaaS itself (e.g., sandbox code execution, permissioned db read/write, web search via Exa, web crawl via Firecrawl, etc.). You can connect custom tools or MCP servers to any agent or orchestrations to supplement its set of tools.

@hunter_powabase  @aiswarya_s Thanks for the upvote! Bring your API key is an option that many of our users prefer. We are also giving our users the option to use powabase's key out of box so they don't have to get their own - should be out in the next product update.

@hunter_powabase  @aiswarya_s Yes, for now — bring an LLM API key (OpenAI, Anthropic, Google, or any model via OpenRouter) and you're set. Embeddings, rerankers, OCR, web search, and scraping are all wired in with no extra accounts. We're actually rolling out an even simpler onboarding soon where new users can start with Powabase's LLM access out of the box (bring-your-own-key stays supported and gets you the best margin). What are you planning to build?

@hunter_powabase  @aiswarya_s hey, quick update, we just shipped an update: bringing your own LLM API key is now optional. New sign-ups get $20 in Powabase credits to start with, covering LLM access (Claude, GPT, Gemini, etc.) along with everything else. Bring-your-own-key stays supported and gives you the best margin at scale, but now you can start building with zero outside accounts.

@aarav_pittman Indeed. A significant time sink during Powabase's design process was figuring out which available tools are best to include. There are so many out there and some optimize for performance while disregarding cost, some try to strike a good balance and aim at best value, and others make the pure open source free-to-use play.

Example: VLM-based text extraction include LightOnOCR-2, Mistral, or PaddleOCR. This is the default approach for our RAG ETL pipeline. But if you know the source is all native text, you can opt into faster and cheaper non-VLM parsers like OpenDataLoader, PyMuPDF (fitz), or pdfplumber.

We generally try to give developers a choice while setting default options to what we believe is the best value for most common use cases.

@aarav_pittman Yes! I've been building apps using Powabase as a backend, and it simplifies the management of all the different tools/integrations needed for an AI app

@odeth_negapatan1 Thanks so much, really appreciate the encouragement!🙏

Great question, it's a classic trade-off: abstraction (convenience) vs. control. Powabase tries to strike the best balance here, informed by our own experience building across a lot of use cases during our dev shop days.

If you're building AI-native apps with RAG or ReAct agent orchestrations, you'll find Powabase easy to use right out of the box. But if your application needs something custom, you're not locked into the full abstractions. You can take the relevant "intermediaries" and build a custom layer on top of them to fit your design specs.

For example, say you need to supplement retrieved RAG context with metadata stored elsewhere. You can invoke the context_handler object to grab the content that normally would have been fed into the agent, then enrich it with your own logic before forwarding it along. Since everything is stateful, you have access and control over all intermediaries by directly handling the databases.

So rather than locking you in, the goal is to give you sensible defaults when you want them. It also saves you tons of coding agent tokens from having to build these abstractions from first principles or gluing them from open source frameworks.

If you share more about what you're building, we'd be happy to advise on which features would be most relevant!

@studentzuo Thanks for asking!

Traces are first-class objects. Everything is API-accessible and persisted in your project's Postgres (ai schema), so you can query, replay, or export it however you like.

You can find details on our documentations page here: https://docs.powabase.ai/concepts/platform-overview. But in general:

Retrieval traces — GET /api/sessions/{id}/runs/{run_id}/retrieved-context returns every chunk the agent saw for that turn with retrieval_score, reranker_score, fused score, source/KB IDs, indexing strategy, and an included_in_context flag. So you can separate what was retrieved vs. what survived reranking vs. what actually made it into context.

Tool-call logs — GET /api/agents/runs/{run_id} returns the full run object: input_messages, output_messages, tool_calls, steps, events, reasoning_steps, usage, plus parent_orchestration_run_id / parent_workflow_execution_id to stitch delegated and workflow-embedded runs back to a parent. Same data streams live as discrete SSE events during execution.

Permission boundaries — Two layers. At the tool level, DB tools take a config_override.schemas whitelist (system schemas rejected, WHERE required for writes), http_request has SSRF validation, and a PreToolUse approval hook can pause execution until you (human) call /approve. At the project level, each project is an isolated stack, and the API enforces Service Role (bypasses RLS) vs. Anon key (respects RLS). Cross-tenant reads return 404, not 403, to avoid existence leaks.

@asherraph Genuinely appreciate this. The tradeoffs you're naming are real, and I'd rather engage them than try to talk you out of your preferences.

BYO Postgres is closer than it looks. Each project's Postgres is a normal Postgres — you get a direct connection string, your app tables live in the public (or your self-defined) schema (the AI features use a separate ai schema), and the self-hosted distribution runs as Docker Compose or Helm on your own infra. In self-hosted mode the "data container" literally is your Postgres. The lock-in surface is the ai schema and the service worker that maintains it — not the database underneath.

The architecture you described is actually a supported pattern. Agents are reachable via API and MCP, and they call tools over HTTP or MCP. So one common setup: keep your orchestration in n8n / Flowise / whatever, and use Powabase as the centralized brain — knowledge bases, sessions, agent configs, and memory live there, but they're invoked from your existing workflows. The inverse works too: Powabase agents can call your n8n flows as custom tools. We tried to keep the layers composable rather than mandatory, so you can adopt one piece without swallowing the rest.

@ozandag Yup, the agent orchestration we implemented follows Claude Code's architecture. We have many built-in tools with detailed configurations and permissions. Some of them rely on third party service providers like Exa and Firecrawl, but your Powabase credits cover their usage. You can also attach your own custom tools via API or MCP.

All agent conversations are put in "sessions" so agents working in the same session can maintain memory. You have full visibility and control over database tables handling sessions and respective agent / orchestration runs. You can also see retrieved context and tool use / output if your agent uses RAG or tools.

In general, there is no need to wire up anything yourself.

If you have specific abstractions in mind that we don't currently support, please share them with me. We're always happy to build it in.

@nurik_shurik Thanks for reaching out!

Really appreciate the depth of this question — it cuts to something we think about a lot.

Powabase's current model is layered defense, not a single identity primitive. The lower layers are designed to compose cleanly with attestation frameworks like KYA at the upper layers.

Where identity lives today. Agents have no ambient access. Every project is a sealed stack — its own Postgres, Kong gateway, GoTrue auth, storage, and AI worker, with no cross-project blast radius. Within a project, an agent's effective rights are the intersection of three things:

1. The schemas and tables explicitly granted when its database tools were assigned (schema-level scoping is configured at tool-assignment time, not at query time).

2. The Postgres role and RLS context of the API key the host app passes in — Service Role bypasses RLS for server-side calls, Anon respects RLS for anything client-reachable.

3. Any hook policies attached to the agent.

Preventing hijacked queries. Even if a prompt injection convinces the LLM to issue something malicious, several hard constraints sit between the model and the database:

• database_query is locked to read-only SELECT, single-statement, results capped, and rolled back after read.

• database_write requires a WHERE clause on every UPDATE/DELETE (no mass writes) and validates all identifiers against injection.

• HTTP/custom tools have SSRF validation and 30-second timeouts.

• The hook system runs at OnRunStart, PreToolUse, PostToolUse, and PreResponse. Hooks can be rule-based (CONTAINS, STARTS_WITH, regex MATCHES, IN), HTTP webhooks that allow/deny/rewrite tool arguments at runtime, or human-approval gates that pause the run on an SSE approval_requested event.

• For chained tools and spawned sub-agents: doom-loop detection terminates the run on 3 identical tool calls, and orchestration recursion is capped at depth 3.

On KYA specifically. We rely on RLS plus the layers above today, not a standalone agent passport. But OnRunStart and PreToolUse HTTP hooks are the natural integration point — a KYA verifier could sit there and attest the agent (and any sub-agent before delegation) on every step, with the hook denying or rewriting the call based on what the passport authorizes. Standardized AI identities are absolutely on our radar as the space matures, and we'd love to do a reference integration. Drop into our Discord if you want to explore that — we'd be interested in trading notes with the KYA team.

@mykola_s The idea came to us about a year ago. We started building by using Supabase's open source codebase as a baseline, then extended it with our own unique features and integrations.

Powabase is not a "static" solution. Given how quickly the AI landscape shifts, we need to be moving with it as well. So the work is never truly finished. Going from Supabase open source baseline to v1 of production-ready Powabase took about 6-7 months.

Shoutout to Claude Code. It's a great coding assistant that really helped us accelerate the development process.

@vrijraj Yes, absolutely. In fact, we already have a good collection of Powabase serving as the backend for a number of projects that we undertook ourselves.

The biggest one to date is a legal research platform made for the Swiss branch of a global insurance company. The goal is to index the entirety of Swiss laws and court decisions to serve as evidence / advisory for eventual claims processing automation.

We are offering a "free MVP build" program where Powabase's forward-deployed engineers build an MVP on top of Powabase for companies with concrete app ideas, for free. You can find out more here: https://powabase.ai/free-mvp/

Would love to hear more about what apps you may be trying to build!

@vrijraj thanks for your support!

@maali_baali Great question! There's a trade off to abstraction level (convenience) vs. control. Powabase definitely tries to strike the best balance based on our own experience building across many use cases during our dev shop days. If your use case involves building AI-native apps with RAG or ReAct agent orchestrations, then you'll find Powabase quite easy to use out of the box.

If your application needs something custom, you don't have to use the complete abstractions offered through Powabase. Instead, you can take relevant "intermediaries" and build a custom layer on top of those to better fit your design specifications. For example, if you need to supplement retrieved RAG context with metadata stored elsewhere, you can invoke the "context_handler" object to get you the content that normally "would have been fed into the agent", then enrich it with custom logic before forwarding it to your agent.

If you share more about what you're trying to build, then we'd be happy to advise you on which features would be most relevant on Powabase!

@maali_baali there's a lot of flexibility built into Powabase for processing data sources and indexing them in different ways. In our experience with building client apps, that's where you can get a lot of value from tuning for your use case, e.g. long vs. short documents, scanned vs. digital native PDFs, images vs text-heavy sources.

@michael_ludden Congrats on the launch, very cool. How do you deal with tool quality/choice paradox for the agents iro selecting the right tool?

@michael_ludden Hey Michael, congrats on the launch. The "4k tools is the headline but selection is the real problem" point upthread is right, and the PageRank-style scoring answer is clever. The follow-up I'd push on is the cold-start and feedback-loop side of that. A PageRank-like system has two known failure modes: a genuinely better new tool with zero agent-reviews stays invisible until someone risks it, and popular tools accumulate usage which accumulates reviews which reinforces rank, so the system can converge on "most-used" rather than "best for this specific task". How does Zero get a brand-new tool its first fair shot, and does scoring stay task-conditional or collapse toward global popularity? For an agent picking among 4k options, "best on average" and "best for what I'm doing right now" can be very different tools. That's the part that decides whether the ranking stays trustworthy as the catalog grows. Following along.

@ketau this!

@debgotwired Thanks Deb! Let us know how we can make you love it even more!

@shreyans_assistiv Thanks for the kind words! Ya we're keeping a close eye on that, with automated systems + human checks ongoing. The agent can only access funds in your wallet, which creates a natural limit. You can also create spending limits within your coding agent.

Would you be interested in spending caps / structures that we put in place? I.e. if your agent spends more than $2 in a minute, we pause all buying until the user approves?

@daniel_baum nice Dan!

@dexter_dethmers2 Thanks Dexter! So glad you're enjoying Zero - lmk if you have any feedback

@michael_ludden  @alwaysunday Thanks Andy! Let us know any feedback - it's still V1 we're open to anything!

@michael_ludden  @aiswarya_s Thanks for the support! The tools are essentially hosted services provided by businesses via MPP, x402, and other agent payment protocols. All you have to do is add the CLI - then you and your agents can discover + use any of the 4000 tools!

@richard_pierre_reid services that require OAuth flow aren't generally among the around 8000 services currently indexed on the platform because of their transactional nature. Did you have something in mind? Thanks for the comment!

@feifanz emdash alert! but thanks Feifan! hehe - much appreciated!!

@daniel_baum  @ryanhudson  @derek_brambles fantastic! can't wait to see what your Brambles AI shopping agents do with it

@othman_katim here's the full list of almost 8000, but TL;DR a crazy diversity of services. web domain hosting, physical postcard sending, image generation, video editing, song creation - all sorts of stuff. check it out! https://www.zero.xyz/browse

@natalia_iankovych here's the full list of services currently indexed by Zero! https://www.zero.xyz/browse

@zeynep_yorulmaz no authentication necessary! transactions only - your agent goes and finds a service, we show the top ranked results based on a bunch of previous agent reviews, runs and other metrics and your agent chooses a service it thinks will do the job you told it to, transacts, and the result is delivered to your agent. transaction done - no auth necessary.

@retain_dev no auth required. check it out and let us know what you think! It's free, heh.

@anand_thakkar1 it's actually closer to 8k now! No auth schemes involved - purely transactional. Tools that require authentication are not part of what Zero is currently solving for because they'd require additional setup steps. Can you give an example of a service you're thinking about btw?

@mythic_dd it's your agent of choice - claude code, codex, gemini cli, hermes, openclaw w/e - so however you have it configured to transact is how it will. I doubt you've set it to "yolo, spend however much you want, go nuts" haha, so you should be good!

@shivam_jaiswal36 Hopefully Zero can help you build better with your agents! Claim the $5, give it a try and lmk how it goes! All feedback welcome.

We help you (and your agents) make informed decisions by scoring each tool, based on agent reviews. Every time your agent uses a tool, it provides a review based on success/failure, value, reliability, accuracy, etc. The scoring then becomes like PageRank.

As usage scales, so will the volume of agent reviews. So the scores become more reliable.

@dhiraj_patel5 no auth involved! handshake between your agent (after you tell it to) and the service provider on the other side for a deliverable. try it out and let us know what you think!

@fberrez1 Thanks for the question! We help you (and your agents) make informed decisions by scoring each tool, based on agent reviews. Every time your agent uses a tool, it provides a review based on success/failure, value, reliability, accuracy, etc. The scoring then becomes like PageRank.

As usage scales, so will the volume of agent reviews. So the scores become more reliable.

Got any ideas to further improve the system?

@tina_chhabra just like any AI agent, you're in control. Zero brings these capabilities to you + your agent - it's your call how the agent engages with it. If you tell your agent to one-shot a solution... it might buy some usage-based tools without approval. If you tell your agent to request approval before it buys, it will.

For any long tasks, Claude shows me Zero capabilities in planning mode. For simple stuff, I just let Claude cook with Zero.

You'll be surprised how far $5 can take you. You agent might spend $0.10 here or there unexpectedly, but you'll learn how to control the agent (just like with any AI system). Give it a try and lmk!

@keesan12 great thoughts and insights. yes, setup friction is real right now, as you might expect for an early, bleeding edge offering like this. and yes we're certainly thinking about a lot of future directions here, but curious what you meant by "We keep seeing runs look productive while they retry the same failure class" - did you mean using Zero? Would love to understand more! Could be very valuable feedback, ty

@binu_george heya! GREAT questions! Gold star haha.

We actually also support MPP in addition to x402, but to answer your question: When an agent hits a 402 Payment Required from a service that's not indexed already by Zero, it is completely up to your agent (or you, depending on how you've instructed it), not Zero, as to whether or not to transact with that service. Zero is a discovery mechanism for such services, as well as facilitation layer for tractions, and generally indexes ones that appear functional, meaning they've successfully run at least one time already and delivered a result that is expected, but if your agent for some reason wants to transact with, say, an x402 service that hasn't yet been indexed by Zero, that would probably put it on Zero's radar to consider indexing. And, to be honest, most services on x402 and PMM are indexed by Zero pretty quickly, unless they're brand new or fail consistently or do something shady.

Micropayment-per-call is definitely already a settlement option for certain API endpoints that people expose via x402 or MPP or other emergent protocols, including ones by Shopify and others, that are compatible with transactional payloads (I pay you I get this thing). Obviously there are some API use cases that can't function in a purely transactional manner, so I wouldn't necessarily say it'll become the only settlement layer for agentic API access - just my opinion here.

The spending cap mechanism currently in place is the instructions you give your agent about how to spend to accomplish a given task, or in general, and the funds you fill their wallet with to use. I personally think of it like an agentic 'allowance' like ya might give a child. ;)

Any follow ups, please ask! Would love your feedback after trying it out.

@kevin_nicholas_chandra let's gooooo! Can't wait to hear about what you build!

@german_merlo1 thanks Germán! Any feedback after using it, please share!

@adamthecreator The privacy-first angle here feels refreshing, especially when so many AI products quietly depend on collecting as much user data as possible. How are you thinking about personalization and “training the browser” while still keeping user interaction data anonymized by default?

@adamthecreator Congratulations to the Kahana team on the launch of Oasis! The privacy-first approach, focus-oriented browsing experience, and vision of making the browser adapt to the user really stand out. Excited to see the journey ahead and wishing the team success on Product Hunt!

@adamthecreator Love the technical ambition behind Oasis. Building a browser that anonymizes interaction data by default solves a massive compliance headache for teams trying to deploy web-scale AI tools without leaking proprietary code.

The core positioning split in the hero section is the only major leak you have right now. Transitioning from a casual consumer vibe like "Relaaax" to complex enterprise links like your "IT & teams guide" creates immediate friction for incoming traffic. If you anchor the core message entirely around isolating data to train local models safely, your activation rates will climb.

Sent over a quick overview of how to structurally tighten that hero sequence to match your B2B targets. Good luck with the launch today.

@aj0671 You know more than anyone how much work has gone into this, and it's an honor to go to war with you today, Atharva. I just wanted to take a moment here and thank you for all that you've done for Oasis, from hooking up wires, to refreshing the most obscure ui and branding assets, to building the next wave of privacy-first ai features in browsing. You're a great engineer, and I'm so proud to work alongside you and have this moment where we can showcase all that we've poured in.

@akansha888 thank you so much for sharing this genuine feedback about Oasis, Akansha. We hate Ad Nauseam, and we are thrilled you feel like Oasis helps you focus and stay aligned. If you have any suggestions for us to double down on this or add new features to help you focus and do deep work in new ways, please let us know! You have no idea how much we appreciation your trying it out, and that you feel this way feels like a miracle to the engineering and product teams :)

@ferdi_sigona Big thanks for raising this important question!

Short answer: Imported site passwords stay on your device, in Firefox’s Login Manager vault (encrypted at rest). They are not part of the local history/bookmark/semantic-index layer, and the Oasis Assistant has no tool or API path to read the vault. Using the assistant still sends prompts and replies (and sometimes page text or tab URL/title) to cloud models—that’s a separate boundary from the password store.

Long answer on how everything works under the hood:

https://www.producthunt.com/p/kahana/oasis-browser-import-bookmarks-passwords-history-and-autofill-data-from-other-browser

@rohan_mehere Woo! We are over-the-moon to hear feedback like this. When we set out to build "Oasis" we chose the name for a reason. We wanted to have a refuge we could genuinely fall in love with, a space where we'd be safe, where we could do deep work while feeling undaunted. If you have any ideas for new ai workflows we can add for you to amplify your capabilities more, please shout them out and we'll get right to shipping away.

@pournami_pottekat that's crazy! The same thing happened to me recently. I had my hands tied while eating a snack and watching TV, and I remembered I could just speak and say "Open Github in a new tab", and then the voice-to-text worked smoothly and Github opened. It was pretty insane lol

@fberrez1  Short answer. Yes, it's not too heavy from what we've seen so far!

Testing with 125 internal beta testers so far, we've received 1 mentions so far of Oasis slowing down performance on a device, but we haven't yet pinpointed whether that's an issue with the application itself or the device/user. We’re still early and tuning performance. If you try it and it feels heavy on your machine, we’d love to know your OS and whether the assistant was open. We’re optimizing for “powerful when you need it, quiet when you don’t.”

This version of Oasis is built on Firefox, so day-to-day browsing experience is in the same ballpark as Firefox.

A bit more detail:

• Browsing & privacy: History, bookmarks, and semantic search indexes stay on device.

• Assistant: Replies are powered by cloud models when you ask Oasis something, so you get capability without keeping a huge local LLM loaded all the time. Though, we have received questions already about supporting local models too

• In practice: If the assistant is idle, resource use should feel close to Firefox. When you’re chatting, using voice, or running browser actions through the assistant, you’ll see the usual spikes (CPU/network) you’d expect from AI + automation.

@hasitha_sigatapu When we first brainstormed the concept of Oasis, before a single line of code was written, we described our design philosophy as "buffing the ice rink." We have tried to focus rigorously on the baseline user experience and preserving the user's focus. We want Oasis to feel smooth as ice, with zero friction. Hearing comments like yours is the ultimate positive reinforcement for us, and it is humbling to feel like we are moving towards are most ancient North Star. While there are other "ai browsers" that may feel like AI has been added for the sake of hype, we have tried to stay true to the idea that Oasis is meant to be an environment for your mind. AI is there to preserve and enhance that environment, but never intrude on your mental flowstate and be the star of the show. The star of the show is always your mental state and capacity to focus and do deep, meaningful work. If you have other constructive feedback about features you'd genuinely want us to add in next, we are chomping at the bit to create new sprints and get back to hacking away :)

@jan_bremec Great use case! Thanks for sharing. Competitor research and doc spelunking is exactly the kind of workflow we had in mind.

Short answer: Today it’s more “help me find and work with what I’ve already seen” than “the browser predicts the next site you’ll want.” It gets more useful as your history, bookmarks, tabs, and hubs fill up locally, but the main lever is what you visited and what those pages were about, not deep tracking of how you scroll or click.

Two layers (worth separating)

1. On your machine — surfacing past work (no cloud model training your profile)

When you ask the assistant things like:

“that Unity doc on rewarded ads I read last week” or “the competitor store page with the gacha breakdown”

It can search:

• Semantic history search — recent browsing (~500 pages) indexed on-device with embeddings built from title, URL, domain/path, and a short text snippet captured when the page was visited. So it’s biased toward meaning (“IAP,” “LTV,” “App Store”) not just “you opened this domain 40 times.”

• Memory search — full-text search across open tabs, tab groups, bookmarks, and history (titles/URLs)

That index grows as you browse (incremental updates, persists across restarts). It does not today behave like a recommender that proactively pushes “you usually open Sensor Tower at 9am.” You invoke it via chat (“find…”, “what did I read about…”) or related tools.

2. How you interact with a page — when you’re on it

That’s a different path: tools like summarize this page or questions grounded in the active tab read visible page content and send an excerpt to the cloud LLM to answer. So for docs/API references, it’s interaction in the moment (you’re on the page + you ask), not the browser silently learning your click patterns over time.

What “trainable” means in Oasis (so expectations match)

• Thumbs up/down + comments on assistant replies → product improvement (anonymous or account-linked; your choice). That’s not on-device fine-tuning of a personal model on your machine.

• Optional “Personalize Oasis Assistant with my account” → ties assistant interaction logs to your account for better signed-in experience over time, still separate from uploading your whole browsing graph.

For your workflow specifically

• Competitor store pages / GDC posts / SDK docs: semantic history + memory search are the wins — “surface that article again” without digging through 200 tabs.

• Long doc sessions: summarize / ask-about-this-page while you’re reading.

• Repeat visits to the same sites: you’ll see them more in search results because they’re in history, but we’re not (today) ranking purely on visit frequency like a dedicated “favorites brain.”

To be super transparent, we’re not yet a full “workflow OS” that learns how you research (e.g. always cross-reference App Annie → spreadsheet → Notion) and automates that sequence without you asking. We’re closer to local recall + assistant actions on tabs/bookmarks/history that compounds the more you use Oasis for that work. Though, we are very interested in expanding our system to handle more complex workflow sequences of any type.

If you decide to experiment, we'd love to know:

1. When you’d want a nudge (start of day, after closing many tabs, when landing on a store URL, never).

2. What the nudge should do (reopen group, search history, summarize, bookmark to a hub — not all of the above).

3. Whether it should be local-only (patterns never leave the machine) vs okay with account-linked suggestions when signed in.

4. One moment last week where proactive would have saved you time — and one moment where it would have annoyed you.
   
   Thanks again for checking us out and supporting the launch! :)

@gkanev Hey Gabriel, thanks for raising these points.

1. Peace of mind comes from the anonymous experience of using Oasis AI. None of your personal data is collected, only raw prompts and outputs which help us create our own "amplifier" model that seeks to improve the speed, accuracy, and quality of commands for you and the broader userbase. I explain this in detail in this video covering our privacy-first approach to training anonymously. You can also see examples of the exact type of JSON payloads that we receive when you send a prompt and learn more about this in our doc on interaction data.

2. We do not support self-hosted models yet, however, that is something we've been exploring and would be open to focusing on in the next sprints on the roadmap. We would also explore allowing you to use your own API keys for other models. Are there any local models you prefer? I've been using Qwen series off Ollama mostly.

3. I'm having trouble understanding your observation here. Which "desktop navigation" or URL are you referring to? Our main website is [https://kahana.co/] if you could point me to an exact page that is not loading, that would be helpful. Or am I missing something?

Thank you again for checking out our campaign and sharing these questions and observations :)

@shivam_jaiswal21 Thanks for sharing your thoughts, Shivam! That shared-profile / sync-leaking-context problem is exactly the kind of thing we had in mind when we separated “your browser” from “how we improve the assistant.”

Short answer: the assistant is not local-only. When you use Oasis Assistant, prompts and replies go to cloud LLMs (routed via our backend/proxy — e.g. Gemini through Supabase Edge). That’s separate from “anonymous training,” which is about whether improvement data is tied to your account, not whether anything leaves your machine.

What stays on your device (by default)

• Browsing history, bookmarks, and the local memory / semantic search index (including on-device history embeddings)

• Saved passwords in Firefox’s encrypted Login Manager — not in assistant memory or training payloads

• Day-to-day browsing profile — we’re not building a “shared browser profile” that syncs your full context to ads or a central browsing graph

What leaves the device when you use the assistant

• Inference: Your message, the model’s reply, and whatever context the agent attaches (e.g. active tab URL/title, tool results, and sometimes page text if you use summarize / page-grounded tools) — that goes to the remote model so the assistant can answer.

• Product improvement (optional logging): With “Personalize Oasis Assistant with my account” turned off (the default for new profiles), we still log assistant interactions to improve the product, but user_id is null — no email, no account block in the payload. “Anonymous” here means not linked to your identity in our DB, not “nothing leaves the device.”

“Anonymous training” specifically

That’s the thumbs up/down + comment flow on replies. You can submit training as anonymous (feedback_events without your user id) or personalized (linked to your account). That’s a second knob from the Privacy setting above — same words, different tables. Either way, you’re giving us signal on what worked; anonymous mode just keeps that row off your account.

Other references: https://www.producthunt.com/p/kahana/oasis-browser-technical-and-interaction-data

Docs: https://kahana.co/docs/technical-and-interaction-data

5-minute YouTube video explaining data that is sent: https://youtu.be/8C3FucA95Lg

What we’re not doing

• We’re not claiming a fully on-device model for chat today (perhaps in the near future!)

• We’re not claiming zero data when the assistant runs — cloud inference + (by default) anonymous interaction logs.

• We’re not uploading your whole browsing history as a training corpus; improvement signal is assistant-shaped (prompts, responses, tab context when you’re in the assistant, tool traces).

For B2B CS workflows: Oasis is closer to “your machine holds the messy browser state; the assistant only sees what you invoke and what tools pull in for that turn” than to “train a shared synced profile.” If you need air-gapped / no cloud at all, we’re not there yet — we’d say that upfront.

Happy to go deeper on any layer (telemetry JSON shape, user-initiated training vs telemetry data from general llm usage, or what’s local for search vs chat).

Just curious, what other aspects are you and your team thinking about? I love this space and these conversations :)

@nlokesh9 Thanks for sharing more about our Journey, Lokesh! The no-ads experience also somewhat snuck up on me. The idea was to always make it add free, but only through sitting in user interviews with Archana and BD, did I realize that no-ads and privacy ranked just as high on the priority list for many users as privacy, anonymity, and data security. We're still learning now with all the responses in this thread!

@pallavi_m18 I hope we can deliver on your dream. What are your dream workflows? What are the series of actions you find yourself doing in the browser? If you could wave a magic wand and have Oasis do anything for you, what spell would you cast? These are the things we'd love to learn more about. When you let us know, we can build in new features that makes it a reality :)

@rashmikadwani it's definitely early, and we feel like we have a wide open frontier to explore. Oasis can be expanded significantly. If you had to say, what could we change or improve? Do you have suggestions? Beyond what you've shared, is there something else we haven't touched on yet that you'd ideally want in Oasis to make your life easier?

@arch14 are there any new types of workflows you’d like us to introduce? Thanks for sharing the positives so far. We still feel like we have a lot of room to improve. If you could wave a magic wand and have Oasis do anything for you, what would you say?

@krishna_samhitha You should try out the "create tab group" and "add this tab to tab group" features! I also switched on "Vertical tabs" by right clicking tabs. I feel like this makes my experience calmer. If you have any wishlist commands holler them. If you had a magic wand and could say anything out loud to manage and reorganize your tabs like magic, what are the things you'd say?

@jaideep_kulkarni it's wonderful to hear that you feel like your experience and workflow have become more structured with Oasis. This is a goal we set out to achieve when we decided to build it, and the fact you feel this way means we are moving in the right direction. We feel like we are just getting started with this, so if you have any wish-list enhancements we can make to the built-in AI, we'd love to know. One that comes to my mind is attacking more complex workflows: if there are more abstract or complex, multi-step workflows you wish you'd be able to do, we'd love to learn more about what those are. We have built a solid foundation and architecture with Oasis AI, and we see a wide open blue ocean ahead to rapidly expand the capabilities of the built-in AI. I'm frothing at the mouth to add more XD

@rashmilamitra we're honored that you've trusted us to give Oasis a try over the last few weeks, and we're delighted to hear that you feel calmer while browsing. I just wanted to emphasize that we are committed to preserving this calm and double down on Oasis as a refuge for you to traverse the Internet and do deep, meaningful work. I'm excited to see what you cultivate and build in your Oasis.

We're also looking for negative feedback of any type. If you ever notice an issue that makes you cringe, whether it's a feature or user experience element, please don't hesitate to give us a shout so we can smooth that out ASAP for you :)

@saksham0403 "Simpler and more natural" has been our North Star from Day 1. That you feel this way makes us feel a degree of fulfillment in achieving our mission. I'm flattered that you're excited about the direction we are taking. While we've tested other ai browsers on the market, yes, we have noticed that some feel like some added another layer of complexity to the user experience, rather than facilitating deep work. These are all positive things, but if you have any constructive criticisms or things you feel like we can improve on for you, please let us know!

@ankit_sai_allam Where would you want it to go Ankit! If you could be greedy, what types of workflows would you dream of in Oasis? What would you want it to automate for you? The more we learn from your perspective and feedback, the better it will get!

@hritik_chalse What could we change or improve? Do you have suggestions? What would you ideally want in Oasis to make your life easier?

@sanku_rajshree_rao I'm so thrilled that our focus resonates with you. Oftentimes, when you're a startup you lack direction, it feels like wandering in the desert. To see so many comments reinforcing that our North Star (user experience and privacy) is a worthy one is quite exhilarating. I would encourage you to not be shy: what do you really want us to do in Oasis? I'd love to understand what people hope and dream to accomplish with it. What outcomes do you want? How should it feel getting there? What does that look like? If we can picture it tangibly we can literally build it!

@sushma_evs Thank you for sharing, Sushma. What about negatives? Any constructive feedback you can provide? Any wishlist features you'd want? We are problem solvers, and we genuinely want your feedback and input so Oasis can reach its potential. We are always looking for ways to improve, and we are committed to rigorously and quickly implementing new features, enhancements, and fixes based on what you tell us.

@niraj_patil3 Cross-Tab Reasoning sounds sick, Niraj! I just added that to the roadmap 🫡

@sonali_ankolikar Don't you worry, Windows version pipeline is being wired up as we speak. I can't wait for you to dive in

@likhitha_guggilla We want to leverage the training data to make the average command in Oasis more fast, accurate, and "quality" in the eyes of the user. If a user executes 100 commands, I want them to feel like it was fast, accurate, and "quality" on 100/100.

@preetika_singh1 Are there any features in particular that you would wish for the most? Really looking for new ideas and a better understanding of how it could enhance your workflows

@jim_jeffers Here's an example of an anonymous payload from training. It's more than thumbs up/down. The user can select a category (badge) and they must enter at least 30 characters of their own words about what was good or bad.

Does this address your question a bit more? I would genuinely love to know what you think.

{
  "badges": [
    "Helpful",
    "Fast"
  ],
  "comment": "It gave me a good response and a good recipe",
  "sentiment": "up",
  "contact_me": false,
  "user_prompt": "can you give me a recipie to make a chocolate cake",
  "training_mode": "anonymous",
  "assistant_reply": "Of course! Here's a classic, delicious chocolate cake recipe that's pretty straightforward to make:\n\n## Classic Moist Chocolate Cake\n\nThis recipe makes a rich, moist chocolate cake that's perfect for any occasion.\n\n### Ingredients:\n\n**For the Cake:**\n*   2 cups (250g) all-purpose flour\n*   2 cups (400g) granulated sugar\n*   ¾ cup (65g) unsweetened cocoa powder (Dutch-processed is great for a darker color)\n*   2 teaspoons baking soda\n*   1 teaspoon baking powder\n*   1 teaspoon salt\n*   1 cup (240ml) buttermilk (or 1 cup milk + 1 tbsp white vinegar/lemon juice, let sit 5 mins)\n*   ½ cup (120ml) vegetable oil\n*   2 large eggs\n*   1 teaspoon vanilla extract\n*   1 cup (240ml) boiling water (or hot strong coffee for deeper flavor)\n\n**For the Chocolate Buttercream Frosting (Optional but Recommended!):**\n*   1 cup (226g) unsalted butter, softened\n*   3-4 cups (360-480g) powdered sugar, sifted\n*   ¾ cup (65g) unsweetened cocoa powder, sifted\n*   ½ cup (120ml) heavy cream or milk\n*   1 teaspoon vanilla extract\n*   Pinch of salt\n\n### Instructions:\n\n**For the Cake:**\n1.  **Preheat & Prep:** Preheat your oven to 350°F (175°C). Grease and flour two 9-inch round cake pans, or line with parchment paper.\n2.  **Combine Dry Ingredients:** In a large bowl, whisk together the flour, sugar, cocoa powder, baking soda, baking powder, and salt.\n3.  **Add Wet Ingredients (except boiling water):** Add the buttermilk, oil, eggs, and vanilla extract to the dry ingredients. Beat with an electric mixer on medium speed for about 2 minutes, until well combined.\n4.  **Add Boiling Water:** Carefully pour in the boiling water (or hot coffee) and mix on low speed until just combined. The batter will be thin – this is normal!\n5.  **Bake:** Divide the batter evenly between the prepared cake pans. Bake for 30-35 minutes, or until a wooden skewer inserted into the center comes out clean.\n6.  **Cool:** Let the cakes cool in the pans for 10-15 minutes before inverting them onto a wire rack to cool completely.\n\n**For the Chocolate Buttercream Frosting:**\n1.  **Cream Butter:** In a large bowl, beat the softened butter with an electric mixer until light and fluffy.\n2.  **Add Dry Ingredients:** Gradually add the sifted powdered sugar and cocoa powder, beating until smooth. Start on low speed to avoid a cloud of sugar!\n3.  **Add Liquids:** Pour in the heavy cream (or milk) and vanilla extract, and add a pinch of salt. Beat on medium-high speed until the frosting is light, fluffy, and spreadable. If it's too thick, add a little more cream; if too thin, add more powdered sugar.\n\n**Assemble the Cake:**\n1.  Once the cakes are completely cool, place one layer on your serving plate or cake stand.\n2.  Spread a generous layer of frosting over the top.\n3.  Carefully place the second cake layer on top.\n4.  Frost the top and sides of the entire cake.\n\nEnjoy your delicious homemade chocolate cake!\n",
  "include_context": true
}

@surabhi_minocha Hey Surabhi! We're still learning all the time right now. The behaviors are developing in real-time. I think we're seeing people enjoy treating it like "Jarvis" from the Iron Man movies. I found myself up late half-watching TV and snacking while I was doing some vibe coding. Then I needed to check on a deployment in Github. Rather than click around, I just realized I could say "Hey Oasis can you open Github workflow tab?"

And then it did it.

After that, I said "Can you summarize the results of the workflow?"

And it did that!

For me and others, we are finding that the ability to use natural language and voice to control the browser lets us stay focused more and reduce mental context switching and cognitive load.

I'm not sure what this would qualify as, I suppose a combination of all of it!

@rushyanth_nerellakunta It sure has been a journey, I'm just in awe of how the team and product have come together. I'll always remember this launch. And the crazy thing is that I feel like we're really just getting started and scratching the surface of what's possible here. So much of the build has architecture and foundations. We're actually getting into spicy stuff now

@alex_calder is the OM2 knowledge graph built on a specific graph database or vector embeddings entirely custom under the hood?

@alex_calder Hey Alex, congrats on the launch. The "$500K/year in December, $15M in May" framing is the realest line here, that curve is exactly why routing stopped being optional. Reading the thread, everyone's circling how to make the routing decision (confidence bands, classifier signals), but Dhruv named the harder one and moved past it: knowing after the fact whether you under-routed. That's the part I'd dig into. A wrong downgrade often isn't visibly wrong, the cheap model returns a plausible answer that's just quietly worse, and in production there's no ground-truth label to catch it. So how does Coworker close that loop? Is there a signal that flags "this got routed cheap and the output degraded," or does it rely on the user noticing and hitting rerun? Because routing quality you can't measure post-hoc slowly drifts, and the savings number stays great right up until trust erodes. That measurement loop seems like the real moat, harder to build than the classifier itself. Following along.

@alex_calder Congrats on the launch. That routing engine simulation in the hero visual is incredibly clean.

Dropped a vote, but looking under the hood at your schema data, there is a serious positioning mismatch on the homepage right now that is likely leaking enterprise conversions during this launch spike.

Your backend schema and enterprise badges position Coworker as a high-leverage, autonomous workflow engine for the GTM stack. You are solving massive operational drag across Salesforce, Jira, and Slack.

But your actual live hero headline ("Same frontier AI. 5x more tokens.") and subheadline lead entirely with an API cost-arbitrage play.

By prioritizing token counts and raw utility pricing, you accidentally reframe an enterprise powerhouse into a cheap developer tool for cost-sensitive engineers. High-ticket enterprise buyers do not optimize for token volume. They optimize for workflow velocity and structural efficiency.

Flipping that hero narrative from a "cheap infrastructure" message to "autonomous operational leverage" will align this massive launch traffic directly with the corporate budget holders.

Insane product build here. Good luck with the rest of launch day.

@benln thank you for your support!

@benln Thank you Ben!!

@benln Thank you!

@kohnigel Congrats on the launch Nigel, looks amazing!

@kohnigel Sick product demo! Amazing and scary how productive people will be using this tech. Congrats on the launch Nigel!

@fberrez1 our 'cheap' primary agent is still a frontier open model. This means that the baseline quality and judgment is still incredibly high and any cheaper subagents are handled by this advisor. And very low latency added. Try it out and let me know thoughts!

@roman_burdyga thanks, appreciate it 🙏

@roman_burdyga thank you!

@anand_thakkar1 yeah, hand-routing gets brutal as the taxonomy drifts. We use confidence bands instead of one cutoff, anything ambiguous defaults up. A wrong downgrade is way more visible than a wrong upgrade so we'd rather burn a bit of cost than ship a bad answer. How big is your task list now?

@irina_sumtsova The "5x cheaper" is a cost-per-task number, not a benchmark claim. On SWE-Bench and Terminal-Bench Kimi is basically on par with Opus, and honestly in most real-world use it's pretty close too. Simple tasks the 5x holds (~$0.39 vs $3.59), and even on complex work the gap isn't as dramatic as you'd expect. And since Coworker knows your context, you can decide if and when you actually need Opus over Kimi.

@thekrew absolutely. We have a feature that lets you rerun your query with a different model so you can compare outputs. Helps users build trust in the cheaper models when they see them hold up.

@retain_dev can't share the internals, but token count alone is a weak signal; complexity doesn't correlate with length. The harder problem is the feedback loop: knowing when you under-routed vs just burned cost you didn't need to. What does your pipeline look like?

@dhiraj_patel5 our organizational memory layer classifies the task before a model is even picked. Looks at complexity, error cost, how much context is needed. So you're not downrouting, you're just routing correctly from the start.

@aiswarya_s great question - the organizational memory layer does eliminate redundant work: context from past conversations, decisions, and actions is stored and surfaced so agents (and humans) aren't re-running the same queries or re-deriving the same answers.

yes you do@saurabh_mohan1. let's get you re-set up with Coworker!

@marianna_tymchuk Thank you!

@ravishankarai_official, when it's ambiguous we default up to the more capable model, false downgrades are way more visible to users than false upgrades. And yeah you can override manually, pick the model yourself or rerun with a different one to compare outputs.

@shivam_jaiswal36, our organizational memory continuously ingests from 50+ connectors so it updates as things change in the source, not on a sync schedule. Role changes, deal updates, doc edits, all automatic. Permissions inherit from the source too. What are you building?

agreed @nithin_raju1!

@natalia_iankovych Our primary agent defaults to frontier open models. It has subagents that can delegate up or down to advisors or subagents using a mixture of open or closed models. And we have an inbuilt review mechnism. This allows us to min cost holding quality at frontier or greater. And no, nothing is hosted in China - everything is in the US.

@thamibenjelloun do check us out!

@maria_anosova Thank you so much, means a lot coming from you. 🙏

@ahmed_mousa6 Thank you for your support, appreciate it!! 🙏

@dhiraj_patel5 Thank you Dhiraj! Super thoughtful question! @Octolane has native merge contact capability and Octolane's AI is smart enough to find the difference to act accordingly.

@shivam_jaiswal36 - Appreciate the kind words, Shivam and yes, data rot from manual field updates is exactly the problem we're solving. If the CRM requires reps to do data entry, the data will always be bad. That's just human nature.

To your question: yes. Octolane tracks conversations across multiple email threads and contacts tied to the same deal. So if you're emailing the champion, their VP loops in on a separate thread, and procurement starts a third, all of that gets unified under one deal automatically. No manual linking required.

We think of it less as "auto-detection" and more as the @Octolane AI CRM just understanding what's happening in your pipeline the way a great sales ops person would.

@jared_salois Great question, Octolane does take care of that with 200+ data sources and building singular context engine to do any task end to end for sales processes as an AI CRM.

@thekrew Thank you for sharing that, this is exactly why we built @Octolane!

@wellsriley Congrats on the launch Wells. very cool tool and love the name and visuals

@kyleledbetter It's because they're mostly bot accounts 🤖

@nithin_raju1 yes!

1. Mojito supports common alts for emoji

2. Supports near-matches/typos gracefully

3. It ranks your frequently-used emoji higher in search—so if you use :pouch: (👝) a lot, it'll rank higher than 💩 over time.

4. Mojito is localized into 18 languages... both the UI and the emoji search (eg: in French, you can search :heart: or :cœur: to find ❤️)

@natalia_iankovych macOS for now.

@mynameisyu no more than you normally do 🙂

@busmark_w_nika Everywhere!

Beautiful work @sebastianko. Massive congrats on the launch!

This is going to help immensely with my tech content creation, where most of my value delivery is visual. I've already tried a couple of use cases on your app, and I'm genuinely loving the Code Editor and Gemini Prompt styles, they look polished right out of the box.

The fact that it's free forever is honestly the cherry on top. Thank you for building this. 🙌

Suggestion for the roadmap: it would be amazing to have a Tweet-style preset that renders posts in their actual format. That single addition would unlock a huge use case for content repurposing for creators.

@ravishankarai_official right now it only supports MP4 and WebM (which you can set in the recording configuration). I will explore adding GIF export as well.

@matheusdsantosr_dev that's a good point. Right now, you can specify precise slide values by using arrow keys (but it isnt very intuitive to do that). Will add that feature!

@madalina_barbu thank you!

@swapnull Hey Swapnil, congrats on launch number two. The "Ringing or Session, resolve it up front" detail is the part that shows this was built by someone who's actually been burned mid-integration, getting that wrong late is exactly the expensive kind of wrong. The thread you opened that I'd pull harder on is the 23-point pass catching API drift across four platforms. That's the real maintenance nightmare in SDK-first integration: the skill scaffolds correct code against today's SDK versions, but six platform SDKs drift independently and constantly. Three months out, CometChat ships a breaking change or iOS deprecates a permission API, and the integration the agent wrote is now quietly wrong. So is the skill the thing that re-detects and re-scaffolds when an SDK moves, or is it a point-in-time generator and drift becomes the developer's problem again? Because "working chat in 10 minutes" is the easy promise, "still-working chat in 10 months without a rewrite" is the one that decides whether teams trust generated integrations. Following along.

@swapnull Smart move forcing the Ringing vs Session split upfront. Mixing CallKit/ConnectionService with link-based rooms always causes painful rewrites. That 23-point verification for VoIP push, permissions, and teardown solves the exact native-bridge headaches I’ve debugged across RN and Flutter. Does it auto-handle foreground/background state transitions, or leave it to the host app?

@katherine_munoz Congrats on the launch Katherine, interested to know how you're measuring dog response?

@midori_verity Start lower than you think. The sound should be present in the room but not front-and-center: if you're talking over it without noticing, that's about right.

We have dedicated receiver apps for Apple TV and iPad, so you can run a session through your TV speakers or position an iPad in the room, controlled from your iPhone. Volume on those works exactly as you'd expect: TV remote for Apple TV, iPad volume buttons for the iPad app. If you're on AirPlay, use the speaker itself. iPhone plays straight from the side buttons.

Trust your dog more than the number: if they glance toward the speaker when it starts, drop it a notch. Unbothered and audible across the room, you're in the right range.

Thanks for sharing it with your friends. Would love to hear how it lands and PLEASE send us feedback on any snag you might find.

@chargewhat_team Thank you! That genuinely means a lot. Shelter dogs are one of the populations we think about most. The environment is everything a calm acoustic system was designed for: unpredictable noise, no routine, no familiar presence. We haven't run formal shelter trials yet, but the stress profiles we built around, loud noise, separation, unfamiliar spaces, map directly to what those dogs experience every day.

Please do send it to your mother. If she tries it with any of the dogs and has feedback, we'd love to hear it. Real shelter data would be incredibly useful to us at this stage.

@natalia_iankovych Ha, the TV trick is a classic, and honestly, for some dogs it works well enough. The issue is that TV audio is mixed for human hearing, so depending on what's on, it can be neutral, calming, or genuinely stressful for a dog. A husky watching a nature documentary with high-pitched animal sounds is a different situation from a husky hearing low background dialogue. The couch surviving is the real metric though.

On verification: the honest answer is that we're in beta and relying on owner feedback and post-session behavioural observation right now. The acoustic parameters are built from published veterinary research on canine hearing and stress response, so the foundation is peer-reviewed rather than intuition. But direct physiological verification in real-world conditions is exactly what the Pawse Tag is designed to do: a collar device that measures what the dog actually hears during a session and correlates it with stress indicators like bark events and movement. That's where we move from "the science supports this" to "we measured this dog, in this session, responding this way."

Your husky would be a great test case. Huskies are vocal and opinionated about sound... and many other things😉

@tina_chhabra Thank you! Henry is doing a lot better. He still has his moments, but the difference from where he started is real.

He's the reason every parameter in this product exists. When you're building something for a dog you know well, you notice things that market research would never surface. The product is better because of him. He says hi!

@naimz Thank you for your message Naim! Right now in Beta it's owner feedback: guardians rate how their dog responded after each session, and that data feeds back into the adaptive loop. It's behavioural observation rather than physiological measurement, which is an honest limitation.

The longer answer is the Pawse Tag. It's a collar device we're building that measures what the dog actually hears in real time, detects bark events and stress indicators, and correlates that data directly with session output. That closes the loop properly: instead of asking the owner how they think the dog felt, we get a direct signal from the dog during the session. The app is the foundation. The Tag is where the measurement becomes real.

Here is a demo rendition:

@nithin_raju1 The apps are already on testflight. You are most welcome to try them out:

iphone: https://testflight.apple.com/join/NV448bpz

iPad: https://testflight.apple.com/join/4v6tXZ7q

AppleTV: https://testflight.apple.com/join/mmp4fKtD

@maltepruser Congrats on the launch team. how do you deal with task debt? i.e today we do 10 of 14 tasks but need to start with 11 tomorrow....

Glad you like the idae, @jared_salois! Users can pick from ALL AI models that exist. We sort by providers, most used, and have a bunch of filters that people can pick and choose between :)

@shivam_jaiswal21 we always keep up with the latest models, we have a person in-house that stays on top of this always. Furthermore, we constantly listen to and integrate user feedback, so if there's something missing, we add it!

@ravi_teja_knts Very cool Ravi, congrats on the launch. How are you handling security and permissions on existing postgres?

@marianna_tymchuk Thanks for the support, yeah I hate when the things I already configured has to change because of a random tool.

@roman_burdyga Thanks for the support

@othman_katim The app already has media management page and rich text editing. I will continue to add more refined settings moving forward so you can just toggle whether any specific section can rich text editor or normal text editor. Currently not thinking of plugins.

However, this is open source app, so you can just clone and add whatever features you need for your personal need.

@fberrez1 yes, dealing with database changes is a struggle. So Basebuddy handles it own permission layer. We have role level and user level permissions. So all users do not have the ability to publish content by default, authors only has the ability to check their own posts by default. However, everything is customisable. You can customise both the editor UI and permissions that fit your specific need.

@shivam_jaiswal21 It connects via the Postgres connection string and operates with whatever role/permissions that user has. So just create a user that has full access to the needed tables, that's it.

@dominic_mangino Hi Dominic, congrats on the launch. I'd be interested to know, practically, how you identify what signal matters most and how to tweak teh environment?

@cslazzar just a smart thermostat!

@nithin_raju1 Couldn’t agree more, Nithin! That was the guiding principle behind Netfox — native performance, no Electron, no bloat, just SwiftUI doing what it does best. Thanks for the support! 🙌

@alina_tyslenok_ Thanks so much, Alina! 😄 Ha, the “3 AM smart bulb” scenario is exactly the kind of thing that pushed me to build Netfox — you never know what’s quietly chatting on your network until you actually look. I’m really glad the plain-English security scan resonated — I wanted it to feel accessible, not like reading a sysadmin log. Appreciate the kind words and the support! 🙏

@daniel_a_a This feels like a much more natural direction for AI writing tools. Instead of forcing people into a separate app, it quietly improves the workflows they already live in every day. How are users reacting to the fully local approach compared to cloud-based AI tools that usually have stronger models but weaker privacy?

@daniel_a_a Hi Daniel. I have been using your application for some weeks now and I really like and RECOMMEND it! Cheers

@daniel_a_a Well done! As others have said, this is auto-complete on steroids and has saved me a ton of typing. Thank you and your team for this!

@slipio Thank you for the feedback! Cotypist's paid plans are for those who derive the most value from it. I hope that Cotypist's free plan will be sufficient for about 80 % of users while still providing a great experience.

@slipio i will agree it does not seem cheap but given how much i use it im going ahead to support the continued development of the product. wisprflow is another ive been using though much less and im thinking it would be cool to see them merge!

@slipio Especially since it is LOCAL AI on YOUR machine. He has ZERO running costs that would excuse these prices. I disabled autoupdate, backed up the dmg and blocked all outgoing connections cotypist tries to make. It works fine. Nobody needs to update to a launch version just to be presented with yet another subscription prompt.

@pappatistos very true. The fundamental problem with this app is the developers ambitious pricing. This feedback has been given to him across the board and he has ignored it. Its a mispositioned and mispriced product being upvoted by people who will never buy it.

@michele_barana - Yes, a lifetime option would be great.

@michele_barana I'll add another suggestion for @daniel_a_a : a purchase parity system that adjusts price based on the country of purchase

@tvdster Thank you for the kind words, and for the early support! I have lots of ideas on where to take Cotypist next, and am glad to have you along for the ride!

Also, happy to hear that Cotypist is working well even on an M1 Mac. In terms of your completion key, where is ; located on your keyboard (this varies depending on the keyboard layout)? I personally am using the key above the Tab key, but I could imagine that a rarely-used key in the bottom right corner of the keyboard could also be a great option to hit with your pinky finger.

@s_mcleod Sam, thank you so much for the kind words and the feedback you’ve been providing over the early access period. I still remember your first message about Cotypist from more than 16 months ago; the product has come a long way since then. I appreciate you sticking with Cotypist (even when it had been causing issues for you) and believing in its potential!

@weichsel Hi Thomas, thank you for the praise! "It feels like it reads my mind" has been a recurring theme in early access feedback. Luckily, the actual insights usually still need to come from oneself, so Cotypist won’t replace that anytime soon.

@csatwood Thank you for your support! You are not alone with that sentiment; I also feel like "typing through molasses" when Cotypist is not available, and others have also reported that any new Mac they set up feels broken until they install Cotypist.

@michael_haessig Thank you for your support Michael, I'm glad to hear that Cotypist has earned your trust (and your money)!

@vonhraban Thank you for the kind words! "This should be built into macOS" is the highest form of praise, I really appreciate it.

@mike_huettemann Mike, thank you for the kind words and the support! Yes, it's been a long journey! I'm glad to hear that the amount of attention to detail and polish in Cotypist is noticeable.

@technocrat Thank you for the endorsement, Richard! I appreciate your support throughout the early access period and am glad to hear that the improvements I made have made a difference for you.

Also, thank you for the suggestion with capitalizing "I"! I think capitalizing "I" and other nouns would be a great addition to Cotypist's existing typo autocorrect feature; I've made a note to add it once the dust has settled after the launch. I'll still need to think about whether to do the capitalization automatically or only when you confirm by pressing Tab; so far, I’ve been hesitant to make any edits to one’s writing without confirmation. Food for thought!

@anand_thakkar1 Great observation - that exact disconnect between static architecture docs and operational dashboards is one of the problems we're trying to solve. Right now, Archi-Flow is designed to support simulation-driven traffic visualization, and we're exploring both directions: ingesting live telemetry from cloud/observability providers and allowing teams to define synthetic traffic patterns for design reviews, onboarding, and failure modeling. The goal is flexibility — use real production signals when available, or model scenarios manually when you're designing systems before they exist. Curious, which Datadog metrics would have been most valuable for you to see directly on an architecture map?

@jan_bremec Exactly — that's one of the problems we're trying to solve. Static diagrams drift from reality quickly, and explaining backend systems usually means jumping between architecture docs, dashboards, and screenshots.

Today Archi-Flow already lets you share outside engineering teams.

How sharing works currently:

1. Build your architecture.

2. Configure traffic simulation if needed.

3. Click Share.

4. Archi-Flow generates a resumable share link containing the architecture state.

5. Copy the link into Slack, email, docs, tickets, or wherever your team collaborates.

6. Anyone opening that link can view the architecture directly — they don't need to recreate it themselves.

For broader communication we also support:

• SVG export → presentations, docs, stakeholder reviews
• PDF export → architecture snapshots and reports
• JSON import/export → engineering workflows and moving architectures across environments

The goal is making architecture easier to communicate — whether that's onboarding engineers, stakeholder discussions, incident reviews, or explaining backend flows to people who don't live inside infrastructure dashboards every day.

@kbyatnal  @fmerian Hey fmerian,

Appreciate you hunting Extend. Parse 2.0 is heavy tech, but most AI startups you hunt lose enterprise traction because their interface architecture doesn't command immediate institutional-grade trust.

I run a high-fidelity visual operation scaling AI-native frameworks and complex dashboards on a strict project basis.

If any of the portfolio founders you back need to convert dense product capabilities into category-defining brands, let’s sync.

@new_user___1452026946a93788355af99 the challenge w/multi-table and mixed narrative comes down to reading order. irregular form means sometimes you have to read a whole column first before the next vs going left to right and up to down prescriptively. for reading order, ground truth is how a human would read a doc to extract meaning.

hey @ingvar_borzov great question, standard LLMs are general-purpose and can be quite costly with high latency for doc parsing, esp on docs with those complex components you listed. You also get a lot less config control and relying on prompt engineering is brittle. Our VLMs are fine-tuned to handle specific layout components like tables, forms, handwriting, barcodes, etc. And we layer on an optional agentic OCR loop for especially challenging edge cases.

Here's a benchmark if you're interested in objective measures of performance! https://www.extend.ai/resources/realdocbench