Manufact, a Y Combinator Summer 2025 startup, is building the hosted pipeline that turns a working Model Context Protocol (MCP) server into a ChatGPT Apps Store or Claude Connectors listing — the latter being Claude's own marketplace surface.
MCP has cleared the first hurdle. The open standard for hooking external tools and surfaces into AI assistants, sometimes called the Model Context Protocol, has shipped in enough production code that calling it "experimental" no longer fits. The next hurdle is getting a tool you built into the places where users actually invoke assistants, then keeping it running there — a problem the company's own framing calls uglier than the protocol work that preceded it.
Manufact, a Y Combinator Summer 2025 company (company page), is taking on that gap. Its product is a hosted pipeline that takes a GitHub repo running the open-source mcp-use SDK and outputs a deployable app aimed at the ChatGPT Apps Store and Claude Connectors, the two consumer surfaces where MCP-style tools currently reach end users.
The company's framing puts the missing layer in MCP not at the SDK level but at the infrastructure level: the work that turns a working local server into a published, observable, cross-client application — branch previews, marketplace submission assets, embedded chat surfaces, and an inspector for replaying the JSON-RPC calls that flow when a user actually invokes the tool.
The company describes six surfaces in its product material (manufact.com). MCP Hosting handles production deploy from a connected GitHub repo on every push. Cross-client testing runs the same tool call against ChatGPT, Claude, and other assistants to check that prompts, schemas, and error paths behave identically. Publishing checks audit the output against ChatGPT Apps Store and Claude Connectors requirements. Cloud Inspector traces, replays, and debugs the JSON-RPC traffic the underlying protocol actually moves. Public chat ships an embeddable chat widget, and Analytics tracks usage, latency, and reliability.
For developers, the entry point is the mcp-use SDK, available in TypeScript and Python. It can be installed as a skill inside AI coding agents such as Claude Code, Codex, or Cursor, or scaffolded from a natural-language description. The original Show HN post describes two core abstractions: MCPClient, which auto-detects the underlying transport and runs server streams as background tasks, and MCPAgent, which wraps the client plus a model and system prompt into tool calls that work across providers. Sandboxed execution and dynamic server discovery round out the utilities. The latter is built to keep a model's context window from flooding when many tools are available.
The hosted layer adds what the SDK does not: one-push deploys without Dockerfile archaeology, preview URLs for every branch, custom domains, and the submission assets a marketplace listing actually needs.
Manufact is not pitching the enterprise. There is no sales motion visible on the landing page, no SOC 2 mention, no procurement checklist. The framing targets independent developers and small teams shipping consumer-facing AI apps — the kind of shop that wants to publish a tool, see if it sticks, and iterate without standing up Kubernetes.
The pricing page (manufact.com/pricing) offers four tiers: a Free plan ($0/mo with $5 in monthly credits, 2 projects, 1 team member), a Hobby plan ($25/mo), a Startup plan ($250/mo), and a custom Enterprise tier. Teams betting on Manufact as a long-term platform vendor should read that page carefully before committing a roadmap to it.
Two numbers the founders lead with are worth treating carefully. The mcp-use SDK has, per the YC company page, crossed 7M+ downloads across Python and TypeScript, and the team reports NASA as a user. Both figures are claimed on company-owned pages; the download figure has not been independently corroborated against npm or PyPI download statistics, and the NASA usage has no named program, team, or public reference at the time of writing.
The open-source repository (github.com/mcp-use/mcp-use) and its stargazer list are live and verifiable signals. The repository metadata endpoint is the cleaner place to read stars, contributor count, and release cadence. As of the launch, the repo shows nearly 10,000 GitHub stars and the kind of activity consistent with a library that is being used, but the MCP ecosystem is early enough that growth rate matters more than absolute count.
Three concerns sit underneath the launch.
First, lock-in. Manufact is selling convenience on top of an open SDK, but the deploy pipeline, the cross-client testing harness, and the marketplace submission tooling all sit on Manufact's side of the fence. A team that adopts the full stack and then wants to leave faces a rebuild, not a port.
Second, marketplace gatekeeping. The whole pipeline ends at two front doors: the ChatGPT Apps Store and Claude Connectors. Acceptance policy, discovery, and ranking on both are outside Manufact's control. A tool that passes every check on the platform can still be delisted by the marketplace, and Manufact has no public leverage to push back.
Third, the open question of whether MCP app distribution reaches enough end users to justify a platform layer above it. The protocol has buy-in from the major labs, but consumer usage of MCP-powered tools inside ChatGPT and Claude is still small relative to the overall assistant user base. If that funnel does not widen, Manufact is selling picks and shovels into a mine that has not yet been dug.
The next month will tell most of the story. Watch the mcp-use repository for release cadence and external contributor growth. That pair is the cleanest read on whether the SDK is gaining trust outside the founders' network. Watch manufact.com/pricing for tier detail, free-versus-paid boundaries, and any usage limits that signal who Manufact expects to convert. And watch the ChatGPT Apps Store and Claude Connectors directories themselves: the number of third-party apps that cite Manufact in their submission metadata is the most direct signal of whether the hosted layer is actually moving tools into users' hands, or just into a waiting room.