MCP·Engineering
← All case studies

mcp-factory — browse the proof before you reply

// keystone proof · public repo · 179 passing tests · 90-second Loom demo
Python + Node 179 passing tests manifest-driven scoped auth fail-soft

The problem

Wiring an AI agent to one tool is a quick SDK wrapper. Wiring it to a dozen tools across several services — each needing scoped credentials, clean error handling, and a registry that doesn't collide — is where projects stall. And when a hiring team is evaluating a contractor, they have no way to verify any of that discipline from a pitch. They have to take your word for it.

What I built

mcp-factory is the manifest-driven engine behind the MCP Integration Sprint. You write one mcp.yaml for a repo and the factory generates the server stub and the ~/.claude.json entry; run the hub and it serves every tool through a single MCP endpoint. The SDK wrapper is the easy part — the engineering is everything around it:

  • Validated, env-scoped manifests — secrets come from the environment, not source. mcp_factory/manifest.py
  • Fail-soft subprocess proxying — a backend going down returns a clean error, not a crashed session. mcp_factory/runtime/subprocess_adapter.py
  • Collision-safe registry — multiple bots' tools coexist behind one endpoint without clobbering each other. mcp_factory/runtime/registry.py
  • A real test suite — green on a clean checkout, not just on my machine. tests/

Evidence you can check yourself

The repo is public on purpose — so you can verify the discipline instead of taking my word for it. Every claim maps to a file you can open:

The README states the number plainly: 179 is the real, reproducible count. Run the suite and you get exactly that — no rounding up, no counting skipped tests as passing.

What it shows about how I work

Proof you can check before you reply, not a polished reel. I build the production layer — auth-scoping, fail-soft, validated config, collision-safety, tests — and I publish it so the discipline is verifiable, not asserted. That's the same engine and the same standard behind a client Sprint.

Book a scoping call  Next case study →