mcp connectors - A tool that builds and verifies MCP servers for connecting to internal company systems
Translated by AIView original
Screenshots
Screenshot 1 of 4
Why I built it
When you use AI agents at work, you eventually have to connect them to internal systems.
Code repos, RDB, DW, metadata catalogs, internal APIs, all of it.
I knew MCP is the standard for that connection, but when I actually sat down to start, every server had its own structure, and I had to think through auth, logging, PII masking and other security handling from scratch every time.
So I made an open source monorepo that builds, verifies and runs MCP servers for internal systems the same way every time.
Key features
- Make servers from a spec: A Claude Code skill interviews you to write a spec, then generates the server, tests and scenarios from that spec
- Shared guardrails: API Key auth, CORS, PII masking, a trust boundary for external text, retries and audit logs, shared by every server
- Inspector and Chat: Call tools one at a time in a web UI, and watch a real LLM like Claude or GPT use the tools through a Trace
- NL2SQL reference: Runs the SQL an agent wrote behind guardrails like read-only enforcement, row limits and blocking cartesian joins
docker compose up and you can run it right away without an LLM key.
License is Apache-2.0.
How it was made
- I set the rules first
I started by writing the coding conventions, directory structure and doc writing rules into CLAUDE.md. Keeping Claude working by the same rules across sessions had the biggest effect.
- I split the plan into steps
I wrote the PRD and work plan first, then split the overhaul into Phases 1 through 6 and went through them one at a time. Each phase got its own branch, and when it was done I logged the verification and merged.
- I split roles across subagents
I set up separate agents for the backend (Python, pytest), Next.js pages, UI markup (shadcn) and code review. When implementation is done, the review agent checks once more.
- I turned the repetitive work into skills
I made writing MCP server specs (/mcp-spec-writer) and building servers (/mcp-server-builder) into Claude Code skills and put them in the repo. The goal was for the next developer to be able to add a server in 30 minutes the same way.
- I automated verification
I set up pytest, vitest, Playwright e2e and CLI scenario tests, and had Claude run them every time it changed code.
Looking back, writing down the rules and the plan in docs mattered much more than writing good prompts.
- Built with
Comments 0
No comments yet
Be the first to comment.