For decades, software companies designed their products for a single type of customer: a human being staring at a screen. Every button, menu, and dashboard existed to translate a person’s intention into a machine’s action. But a small startup based in San Francisco and Zurich believes that era is ending — and that the future belongs to companies that build software not for people, but for the artificial intelligence agents that increasingly act on their behalf.
Manufact, a three-person company that emerged from Y Combinator’s Summer 2025 batch, announced in February that it raised $6.3 million in seed funding led by Peak XV, the venture capital firm formerly known as Sequoia Capital India and Southeast Asia, which now manages more than $10 billion in assets. Liquid 2 Ventures, Ritual Capital, Pioneer Fund, and Y Combinator also participated in the round, alongside angel investors including the co-founder and chief operating officer of Supabase.
The company’s thesis is deceptively simple and potentially enormous: as AI agents take over more of the work that humans perform inside software applications — filing expense reports, managing customer support tickets, writing code, booking travel — every software product on earth will need a new kind of interface designed specifically for those agents. Manufact is building the open-source tools and cloud infrastructure to make that transition possible.
“Software products are already being accessed by and will be accessed mainly by AI agents, or by users through chat interfaces,” Luigi Pederzani, co-founder and co-CEO of Manufact, said in an interview with VentureBeat. “That’s our bet. That’s our thesis. And that’s what we are really rooting our company on.”
How Anthropic’s Model Context Protocol became the universal standard for AI agents
To understand Manufact, you first have to understand the technology it is built on: the Model Context Protocol, or MCP, an open standard introduced by Anthropic in late 2024 that has rapidly become the dominant way for AI agents to communicate with external software tools and data sources.
Before MCP, connecting an AI agent to a company’s software required custom integration work for every single tool — a bespoke connector for Slack, another for Salesforce, another for a database. It was tedious, expensive, and fragile. MCP standardized this process into a single protocol, functioning as what CIO magazine recently called “the USB-C of AI” — a universal connector that lets any AI model plug into any software system through a single, consistent interface.
The adoption has been explosive. In December 2025, Anthropic donated MCP to the Linux Foundation’s new Agentic AI Foundation, co-founded with Block and OpenAI, with support from Google, Microsoft, Amazon Web Services, and Cloudflare. More than 10,000 active public MCP servers now operate across the ecosystem. ChatGPT, Cursor, Google Gemini, Microsoft Copilot, and Visual Studio Code all support the protocol. Enterprise-grade deployment infrastructure exists from AWS, Cloudflare, Google Cloud, and Microsoft Azure. An estimated 7 million downloads of MCP servers occur every month.
“Great protocols are as good as their adoption,” Pederzani said, drawing a comparison to the mobile revolution. “We saw the same transition with mobile, right? In the beginning, companies were just creating a pretty simple mobile app. Who would have bought a hotel or a flight or used a bank account from a mobile app? But as time passed, the web became mobile first. What we think is that software products will be MCP first, or chat first.”
The stakes are high. The global AI agents market reached $7.84 billion in 2025 and is projected to surge to $52.62 billion by 2030, according to industry analysts. The MCP Dev Summit, the largest conference dedicated to the protocol, takes place April 2–3 in New York City under the Linux Foundation’s banner, with speakers from Docker, Workato, and major cloud providers — and Manufact will be among the companies presenting.
Two Italian founders, a Zurich co-working space, and an open-source library that went viral
Manufact’s origin story reads like a case study in the power of open-source communities to validate a startup idea before a single dollar of venture capital is raised.
Pietro Zullo and Luigi Pederzani, both originally from Italy, met at a co-working space in Zurich — the same space that produced Browser Use, Bloom, and other startups that went through YC in previous batches. Zullo was studying at ETH Zurich; Pederzani was working at Morgen, an ETH spin-off AI startup used by teams at Spotify, GitHub, and Linear, after leading a 12-engineer team at Accenture Switzerland. Both were winding down previous projects in early 2025 when MCP launched.
“We both wrote agents in the past, and it was such a mess to write the tools, the integrations,” Zullo recalled. “When MCP came out, it looked like the perfect fit for what we were trying to do. But only Cursor, Claude Code, a few closed-source applications allowed you to actually use the protocol. I don’t think I’m going to do groceries or browse the internet or check my emails from Cursor — it’s like, not the right code, right? So we wrote an open-source library to basically do what you could do in Cursor with MCP servers, but on your own machine, on your own application, in your own terms.”
They called the library mcp-use, with a slogan that resonated across the developer community: “Connect any MCP to any LLM in six lines of code.” The repository attracted 2,000 to 2,500 GitHub stars within weeks. Today, the SDK has surpassed 5 million downloads and 9,000 GitHub stars. Organizations including NASA, Nvidia, and SAP use the library, and Manufact claims that 20 percent of the US 500 have experimented with it.
“The amount of power that you can put in six lines of code was really staggering,” Zullo said. The pair applied to Y Combinator on the day of the deadline. “We were super spontaneous because we had this open-source vibe and just enjoyed the process. We had so much energy from the community that was lifting us up, and we knew it was going to be fine.”
Inside Manufact’s plan to become the ‘Vercel for MCP’ — from SDK to cloud in 60 seconds
Manufact’s strategy borrows directly from the playbook that turned Vercel into a multi-billion-dollar company by providing hosting and developer tools for front-end web applications. The analogy is deliberate: just as Vercel made it trivially easy to deploy a Next.js app, Manufact wants to make it trivially easy to build, test, and deploy the MCP servers and MCP apps that AI agents need to interact with software.
The company offers three core products. First, the open-source mcp-use SDK, available in both Python and TypeScript, lets developers spin up a fully functional AI agent connected to MCP tools in as few as six lines of code. It supports any large language model, including local models, and has integrations with LangChain and other popular frameworks. Second, a built-in inspector and testing suite allows developers to visually debug their MCP servers in a browser, view raw JSON-RPC traffic, and test tool execution in a sandbox — without connecting to a live AI agent. Third, the Manufact Cloud platform handles deployment, scaling, authentication, access control, and observability, allowing teams to go from a GitHub push to a production MCP server in under 60 seconds.
“As software becomes more agentic, the hard part isn’t the model anymore — it’s everything around it,” Zullo said. “We started Manufact because developers were spending too much time on plumbing instead of building and shipping their products.”
The company has also moved aggressively into MCP apps, a newer extension of the protocol that allows developers to render interactive user interface components — React widgets, data visualizations, input forms — directly inside chat clients like ChatGPT and Claude. Manufact’s SDK lets a developer scaffold an MCP app with a single terminal command, edit React widgets, and deploy to ChatGPT in under a minute. This positions the company at the center of a potentially massive new distribution channel: ChatGPT alone has more than 800 million users.
5 million downloads, zero revenue, and a crowded field of cloud giants
Every open-source company faces the same fundamental tension: the community that makes the project valuable is not the same thing as a paying customer base. Manufact has been candid about this challenge.
Pederzani said the company made a deliberate decision after Y Combinator to focus entirely on the open-source product and community, rather than rushing to monetize. “A lot of open-source projects jump immediately on the monetization part and kind of betray the community,” he said. While NASA, Nvidia, and other prominent organizations use the SDK, Pederzani acknowledged they are not paying customers. Manufact’s target is to reach $2 million to $3 million in annual recurring revenue by the end of 2026, which would position it for a Series A fundraise.
The competitive landscape is crowding fast. AWS, Cloudflare, Vercel, and Docker have all launched MCP hosting features. But Manufact’s founders argue they sit in a complementary position relative to the model providers. “Anthropic and OpenAI are betting that their own chat products — Claude and ChatGPT — will become the primary interfaces through which people access all software,” Pederzani said. “If that bet plays out, we will serve these systems. That’s going to be massive.”
Why software companies without MCP servers risk becoming “dumb databases” for AI agents
Behind Manufact’s optimism lies a darker observation about the software industry that gives their pitch urgency. Pederzani argued that companies that fail to make their products accessible to AI agents risk being reduced to “systems of record” — dumb databases that agents query but that no longer own the user experience or the customer relationship.
“Now we have customers that come to us and say that their customers are choosing to adopt their product over a competitor because they offer an MCP server,” Pederzani said. “At the same time, there is a threat here that could put companies to become just systems of records. And this is really something that a lot of companies are scared of.”
In late February, Manufact co-hosted what it called the largest MCP apps hackathon to date at Y Combinator’s headquarters in San Francisco. The event drew 650 applications and 300 builders. OpenAI, Cloudflare, and Anthropic all sponsored it. Perhaps the most telling detail: eight employees from Anthropic attended — more people than Manufact’s own three-person team. The model providers, it appears, view Manufact as an ally rather than a threat.
Three employees, $6.3 million, and the ambition to capture a share of every AI tool call on Earth
For all its momentum, Manufact faces significant headwinds. The company has just three employees and has not yet demonstrated a scalable revenue model. Its most high-profile users are not paying customers. The $6.3 million seed round provides limited runway in an industry where infrastructure companies often require substantial capital to reach profitability. And the cloud providers that have launched MCP hosting features already own the customer relationships and billing infrastructure that enterprise buyers rely on.
But when asked what success looks like in two years, both founders pointed to a single metric: the percentage of global AI tool calls that flow through their infrastructure. “Our metric is the global tool calls or servers that run on Manufact — how many tool calls are passing through Manufact, made by agents,” Pederzani said. “Like Stripe is doing for the global GDP. We’re going to win if we can get a great number for it.”
The Stripe analogy is ambitious — Stripe processes hundreds of billions of dollars annually and is valued at roughly $90 billion — but it captures the scope of what Manufact’s founders believe is at stake. If MCP becomes the universal standard through which AI agents interact with all software, the company that provides the infrastructure for building and deploying MCP servers could occupy a position of outsized influence.
“In the end, what matters is to make something agents want,” Zullo said, riffing on Y Combinator’s famous dictum to “make something people want.” “What we’re focusing on and what we’re building is to help this transition of building for agents instead of building for humans.”
[/gpt3]
