Introducing mcpfold

MCP config has a sprawl problem. Every client stores its servers differently — Cursor and Claude use mcpServers, VS Code uses servers with ${input:} secrets, Zed uses context_servers, Windsurf can't even call authenticated remotes natively. So you copy-paste the same servers into five files, hardcode tokens into plaintext JSON, and every server you add dumps its full tool schema into your agent's context on every turn — used or not.

mcpfold keeps one canonical mcp.config.jsonc and folds it out to every client, each in its own format. Secrets are references (${env:…}), never values, so nothing sensitive lands on disk. And a local proxy curates each server's toolset per client, so connecting everything doesn't tax your context window.

The wedge

npx mcpfold init      # scaffold your canonical config
npx mcpfold import    # adopt the configs you already have
npx mcpfold sync      # fold it out to every client

diff gates drift in CI, doctor catches the silent footguns, and sync --watch keeps every client current as a background habit. It's MIT-licensed and runs entirely on your machine — no account required.

Why now

Agents are only as useful as the tools they can reach, and the client ecosystem is fragmenting fast. Breadth is the moat: mcpfold folds to eight clients today and adding one is a one-PR job. Read the context-window benchmark for the numbers, or just install it and fold your first config in under a minute.