Manufact Has 5 Million Downloads and NASA on Its Client List. Here Is Why MCP Infrastructure Matters.

The Macro: MCP Is the USB Port for AI Agents

I will make a prediction that sounds obvious in retrospect: the companies that build the infrastructure for AI agent connectivity will be worth more than most of the companies building the agents themselves.

Here is why. Every AI agent needs to interact with external tools, databases, APIs, and services. Right now, most of that integration is custom code. You build a plugin for Slack. You write a connector for your CRM. You hardcode API calls to your database. It works, but it does not scale, and it creates a maintenance nightmare.

The Model Context Protocol changed this. MCP provides a standard way for AI models to discover and use external tools. Think of it like USB for AI. Before USB, every peripheral needed its own proprietary connector. After USB, everything just worked. MCP is doing the same thing for AI agent integrations.

But having a protocol is not enough. You need SDKs that make it easy to build with the protocol. You need hosting infrastructure to deploy MCP servers. You need debugging tools to figure out why things break. You need authentication and access controls for enterprise deployments. You need the entire stack, not just the spec.

This is where Manufact comes in. They started as mcp-use, the open source SDK that became the default way developers build MCP integrations. Five million downloads. Nine thousand GitHub stars. Adoption at 20 percent of Fortune 500 companies. Those are not vanity metrics. That is real infrastructure adoption at scale.

The closest comparison is Docker in its early days. Docker did not invent containers. It made containers easy to use and then built a platform around them. Manufact did not invent MCP. They made MCP easy to use and are now building the platform around it.

The Micro: A YouTube Creator and an Accenture Engineer Walk into a YC Batch

The founding story is unusual. Luigi Pederzani was a founding engineer at Morgen, an AI startup used by teams at Spotify, Linear, and Canva. Before that, he led a 12-engineer team at Accenture Switzerland. He also built a YouTube channel to 150,000 subscribers and 15 million views. That last detail matters because it shows he understands developer marketing, which is half the battle for open source companies.

Pietro Zullo is the other co-founder. The team is three people total, split between San Francisco and Zurich.

They went through Y Combinator Summer 2025 and rebranded from mcp-use to Manufact as they shifted from pure open source to a commercial platform.

The product has three layers. First, the open source SDK that lets you build MCP agents in six lines of Python or TypeScript. Second, a cloud platform for deploying and managing MCP servers, with a built-in inspector for visual debugging and a GitHub App for automatic builds. Third, enterprise features like role-based access controls, agent persistence across sessions, and centralized configuration.

The customer list is legitimately impressive for a three-person company. NASA, NVIDIA, SAP, IBM, Oracle, Red Hat, Elastic, Intuit, and Verizon are all using the platform. The website shows over 4,000 companies using the SDKs and 1.5 million MCP agents executed through the platform.

The product lets you build ChatGPT apps with interactive React components, deploy Claude connectors, and create custom MCP servers for internal tools. The inspector tool is particularly smart because debugging MCP integrations is genuinely painful without good tooling. You are dealing with JSON-RPC messages flying between models and tools, and without visibility into that pipeline, you are debugging blind.

No pricing is listed publicly, which suggests they are doing custom enterprise deals rather than self-serve pricing. That is the right call at this stage. When NASA is your customer, you negotiate contracts, not publish a pricing page.

The Verdict

Manufact is in the rare position of having already won the open source layer. Five million downloads and real enterprise adoption means the SDK is the standard. The question now is whether they can convert that open source momentum into a commercial platform business.

The competitive landscape is getting crowded. Anthropic maintains the MCP spec and could build their own hosting platform. LangChain, CrewAI, and other agent frameworks are adding MCP support. Cloudflare and Vercel could offer MCP server hosting as a feature. The risk for Manufact is the same risk every open source company faces: the protocol becomes so standard that the hosting layer gets commoditized.

But right now, nobody else has the combination of developer adoption, enterprise customers, and tooling depth that Manufact has. They are the Rails of MCP. They did not invent the protocol, but they made it usable, and now they are building the Heroku on top.

Thirty days, I want to see public pricing and a self-serve tier. Sixty days, I want revenue numbers or at least a clear signal that enterprises are paying, not just using the free SDK. Ninety days, the question is whether MCP server hosting becomes a real revenue line or whether the value stays in the open source layer where it is hard to monetize. If Manufact can charge for hosting, observability, and enterprise features the way Vercel charges on top of Next.js, this is a company that could define the AI infrastructure category.