Manufact (formerly mcp-use)

Software products will be MCP-first. Companies that don't build for agents risk becoming just "systems of record." The ones that do will own the next wave. Huge thanks to @MichaelFNunez for taking the time to understand our thesis and tell our story on @VentureBeat with such depth and care. venturebeat.com/infrastructure…

Join us!

@manufact @pietrozullo @OpenAIDevs @mcp_use best observability, the cli experience is super useful and the mcp-use framework makes it incredibly easy to use manufact's powerful features right out of the box. being able to transform from standard mcp to mcp apps is just a matter of a few hours. couldn't be happier!

We're hosting a GLOBAL Generative UI Hackathon with @AITinkerers 🌟 15+ cities, 4 continents, all at once! You get a Generative UI starter kit with all the technologies pre-hooked up with MCP servers + Agent skills. So no time wasted on wiring infra. Includes A2UI, AG-UI and MCP Apps. Each city will follow the same format: a global kickoff video plays at the start (with the co-creators of each protocol), teams then build, and finally a show-and-tell wraps the day. 📆 THIS SATURDAY, May 9th Register now: virtual-events.aitinkerers.org/p/the-generati… 🏆 WINNERS: Mac minis for the entire team 🥈 Runners Up: Meta Raybans And credits + merch for both! Featuring @GoogleDeepMind, @CopilotKit, @Manufact, @LangChain and @daytonaio These kinds of events happen once in a blue moon. And it's ONLY possible with the @AITinkerers community.

5/ Our agents curate market signals 24/7, test ideas, and deliver intelligence directly to yours so they can operate autonomously by querying it. *Delivered via REST API, CLI & MCP (@manufact) There are three core primitives: 1. `investigate` 2. `subscribe` 3. `contribute` And one proactive intelligence layer: 4. `briefing` That is the product language. Everything else supports those four ideas. Investigation An investigation is a scoped market research job. You give MSX a hypothesis or question, like: - “What should I build for solo founders struggling with quarterly taxes?” - “Who should I sell AI bookkeeping automation to?” - “What content gap exists in founder finance?” (Initially based on a list of filters / niches / brands we have data about, that will increase with time) MSX then: - scans the curated signal graph - pulls relevant raw signals - dedupes repeats - clusters them into problem areas - validates whether those problems are actually unmet or already solved - scores urgency and confidence - returns a structured result So an investigation is not “search.” It is closer to: “run a market thesis against a live intelligence layer.” An investigation should return things like: - top problem clusters - opportunity summaries - confidence score - urgency score - citations and source URIs - timestamps - labels like `emerging`, `saturated`, or `noise` Mental model: - search = “find me posts” - investigation = “tell me what matters, why it matters, and whether it is worth acting on” Subscription A subscription is a standing watch. Instead of asking the same question every hour, your agent says: - “Notify me when this market changes” - “Wake me up if urgency crosses 0.9” - “Tell me if a new cluster emerges in this niche” - “Stream deltas for this investigation” So subscription is push, not pull. A subscription should: - attach to a query, hypothesis, niche, cluster, feed, or investigation - define a threshold, trigger, or condition - send only meaningful changes - deliver via webhook, stream, CLI output, or MCP callback flow Examples: - a builder agent subscribes to new demand spikes in bookkeeping - a writer agent subscribes to rising founder pain points - a seller agent subscribes to new ICP segments showing high frustration - a strategist agent subscribes to shifts in saturation or competitive gaps Mental model: - investigation = “tell me what is true now” - subscription = “tell me when truth changes” Briefing A briefing is proactive market intelligence delivered on a schedule, even when the agent does not know what to ask. This is the “weekly newsletter for agents” layer, except structured for action instead of passive reading. A briefing should answer: - what trends are emerging? - what opportunities matter most this week? - what content should I create? - who should I sell to? - what talent or humans should I work with? - what changed since the last briefing? - what should I investigate next? So a briefing is not question-driven like an investigation, and not threshold-driven like a subscription. It is schedule-driven. A briefing should: - be generated from the agent’s profile - rank the most important signals, opportunities, and changes - include summaries, citations, confidence, and urgency - suggest follow-up investigations - be delivered on a cadence like daily or weekly Examples: - a builder agent gets a weekly brief of the best unmet demand clusters - a writer agent gets rising topics and under-covered content angles - a seller agent gets newly promising customer segments - a talent agent gets relevant humans, operators, or experts to hire or work with Mental model: - investigation = “answer my question” - subscription = “alert me when something changes” - briefing = “tell me what matters before I even ask” Contribution A contribution is the write-back loop. Your downstream agent used MSX, acted on it, and learned something in the real world. That result should come back into the system. Examples: - “I shipped this product and got 47 paid users in week 1” - “This signal converted badly” - “This niche had high frustration but low willingness to pay” - “This messaging angle outperformed the others” - “This prospect segment responded” Contribution matters because it turns MSX from a read-only intelligence system into a learning runtime. A contribution should: - include the observation - link back to a cluster, investigation, briefing, or market segment - contain evidence or metadata - respect tenant privacy - improve future ranking, validation, and personalization Mental model: - investigate = perceive - subscribe = stay aware - briefing = stay oriented - contribute = teach the system what happened

openai accepted 150 apps to the @ChatGPTapp directory in the last two weeks. 500 more in the queue. ~1B weekly users on the other side. if you run an MCP server, this is the biggest distribution channel that opened in years. we just wrote down the entire submission flow, screenshots and all, including the silent failure modes that send you back to the start. written for you and your coding agent: drop it in and it'll see every prereq and every blocker before you open the form.

Most companies aren't ready for agents. They've wasted months integrating AI into their product, only to realise AI isn't a feature, it's a user. About 6 months ago, the dominant view of "integrating AI" was to embed it into the product. Add a chat feature. Bolt on a copilot. Fuse AI with the existing UI. That hybrid model has proven to be an awkward middle state. Useful in patches, but not the final form. A clearer model is emerging. Treat AI as a user that consumes your product. Your product becomes a tool. An API, MCP, a CLI. And the agent (Claude, ChatGPT, Claude Code, Codex, OpenClaw or whatever comes next) becomes the entity that operates it on the user's behalf. This is called going headless. The interface layer dissolves. Your product is no longer something a human navigates through a browser; it's a capability surface that an agent invokes from wherever the user already is: Slack, voice, a chat window. Once you accept that framing, almost every assumption about how to build a product changes. Salesforce has now publicly committed to this. Their Headless 360 announcement exposes the entire platform (Salesforce, Agentforce, and Slack) as APIs, MCP, and CLI. In their own framing: "Our API is the UI." But going headless is only half of the shift. The second half is distribution. Claude has just announced organic discovery within their connector ecosystem. Agents can surface and invoke connector apps based on user intent, without the user having to install or even know they exist. Every other major LLM is likely to follow. That changes the strategic equation. It's no longer enough to ship an MCP server to a connector store. You also need to optimise for how agents discover and select tools. Discoverability becomes a first-class concern, in the same way SEO became one for the web. A year from now, most major software companies will have gone headless. The companies winning the next phase of AI aren't the ones integrating it into their products, they're the ones rebuilding their products to be consumed by it. The ones that also figured out agent discovery early will have meaningfully more distribution than those that only solved the architecture. As @manufact say, make something agents want.

customer’s love! ❤️ that’s the whole point!

@yigitkonur @pietrozullo @OpenAIDevs @mcp_use Thanks for your feedback! Would love to learn more about what you like the most :)

@pietrozullo @manufact @OpenAIDevs @mcp_use the best platform for dev exp

big news @manufact was just listed in @OpenAIDevs docs as a recommended provider to host your MCP App, super proud of this and happy that the work we have been putting on it is being recognized if you are developing an MCP App/Server, @manufact cloud will let you ship 10x faster, to start just connect a github repo and see the magic happen more to come soon 👀

@linear Good. If you want to build one you can use Manufact to deploy it.

New: MCP support for Linear Agent Connect tools like Granola, Notion, PostHog, and many others to pull in context for creating issues, drafting PRDs, and writing project updates.

Monitor the situation. MCP Server. Open source. Use it on ChatGPT / Claude. Built with mcp-use and deployed on Manufact!

@pederzh tokens go 🔥🔥🔥

you can vibecode those ChatGPT apps for free with manufact.com/vibe

is it me or does tech Twitter not care about MCP Apps???

@pederzh @wesbos @dsp_ deployed on Manufact!

@wesbos @dsp_ we got great results with the remotion MCP App and Claude remotion.mcp.fast/mcp

WHO IS MAING THESE VIDEOS they are so well done. I need to see the process they use to slice up and animate the UI

authentication is the thing that takes the longest to click when you start building MCP in production there are actually two separate problems to solve: 1. how does the AI client (ChatGPT, Claude) authenticate to your MCP server 2. how does your MCP server authenticate to your backend APIs most guides conflate these. we wrote the complete breakdown coming out of a session we had with a 1B+ revenue financial company that handles very sensitive data manufact.com/blog/authentic…

@reallynattu That’s super cool but it’s local! If you want to build an hosted version, count on us!

We just connected Lottie Creator to AI via MCP. You describe the animation. Claude, Codex, Gemini reads your file, edits layers, generates variants, ships dotLottie; without leaving the editor. This is what we built Lottie Creator for. What animation would you build first?

tldr: fork-to-star ratio is the strongest simple heuristic for identifying potential GitHub stars manipulation! - standard: 1.2 - potential manipulation: 0.5 proud of @manufact / @mcp_use fork-to-star ratio of 1.22 (in line with other AI tools like: crewAI, dify, agno, mem0, browser-use) 💀watch out for anything with a 0.5 ratio

👋🏼 we are at the MCP Dev Summit NA in NYC! come say hi at our booth!

That’s why the developer adoption of solutions to deploy MCP servers and apps like @manufact is exploding!

cool merch by @manufact

a chatgpt app that finds the best deals for multi-day stopovers. fly cheaper and visit new places built with @manufact and @skiplagged