mcpfold◆mcpfold.yaml — one source of truth
servers:
github: { tools: [create_issue, get_pr] }
supabase: { tools: [query, insert] }Open source · local-first
mcpfold — one MCP config for every client
mcpfold is a free, open-source CLI that manages your MCP servers from one canonical config and folds it out to all 18 supported MCP clients — Claude Code, Cursor, VS Code, Windsurf, Zed, and more — each in its own format.
Curate each server’s tools to skip the context-window tax — every server dumps its full tool schema into context, used or not. Curating the toolset cuts tool-schema tokens by ~80% (7,476 → 1,497), with secrets kept as ${env:…} references, never hardcoded values.
npx mcpfold init
context-window tax7,476 1,497 tokens~80% smaller
CClaude Code
mcpServers: {
github: {…}
supabase: {…}
}CClaude Desktop
mcpServers: {
github: {…}
supabase: {…}
}CCursor
mcpServers: {
github: {…}
supabase: {…}
}VVS Code
servers: {
github: {…}
supabase: {…}
}WWindsurf
mcpServers: {
github: {…}
supabase: {…}
}ZZed
context_servers: {
github: {…}
supabase: {…}
}Feel the tax — and the savings
Pick your servers and how many tools each agent actually needs. Example servers from the published benchmark — no endorsement implied.
Tools
45 → 9
Tool-schema tokens
7,476 → 1,497
Context cut
80%
Tokens use the 1-token≈4-chars approximation for instant estimates. The benchmark publishes exact GPT and Claude counts that land within a few percent.
Manage every MCP server from one file
Write your MCP servers once in a single canonical mcp.config.jsonc — the source of truth mcpfold folds out to every client. Add, curate, and version your MCP config in one place instead of hand-editing a different file for each tool.
Works with every MCP client
mcpfold renders each client's own native format from one config, so the same MCP servers show up everywhere:
Curate tools, cut the context-window tax
Every MCP server dumps its full tool schema into the model's context whether you use those tools or not. mcpfold lets you allow- or deny-list tools per server, so only what you need loads — the committed benchmark shows the token savings.
Secrets as references, never values
Your MCP config carries ${env:…} / ${op:…} references instead of raw tokens, so nothing sensitive is ever written to a client config or committed to git. mcpfold resolves the reference at fold time, from your environment or secret manager.
How it works
Four commands take you from scattered client configs to one source of truth, folded out everywhere.
mcpfold initCreate the canonical mcp.config.jsonc — one source of truth for every MCP server you run.
mcpfold initmcpfold importPull the servers already configured across your clients into the canonical config, secrets rewritten as references.
mcpfold importmcpfold syncFold the canonical config out to every client's native format — add a server once, it appears everywhere.
mcpfold syncmcpfold diffSee any drift between the canonical config and what each client currently has, before you sync.
mcpfold diff
Who mcpfold is for
Solo developers
Run the same MCP servers in Claude Code, Cursor, and VS Code without hand-editing three config files. Set it up once; every client stays in sync.
For solo developers →Teams
Share one reviewed config, keep secrets as references, and see every change in an audit trail. The hosted cloud syncs it across the team.
For teams →Power users
Curate exactly the tools you want from a large directory of MCP servers, and cut the context-window tax on every model call.
For power users →Free, open source, and MIT-licensed
- GitHub
- Latest on npm: v1.4.1
- MIT license
Explore mcpfold
- Install — npx, npm, Homebrew, or Scoop — no account needed.
- Features — deep dives on the four pillars, with examples and proof.
- MCP server directory — browse and add community MCP servers.
- Client setup guides — add MCP servers to Claude Code, Cursor, VS Code, and more.
- MCP glossary — what an MCP server, client, and the protocol actually are.
- Documentation — config format, secrets, adapters, CLI contract.
- How mcpfold compares — by hand vs a hosted gateway vs mcpfold — honestly.
- Pricing — free CLI (MIT); optional paid cloud for teams.
One config for every MCP client
Install the free, MIT-licensed CLI and fold your MCP servers out to every client — or add the hosted cloud when your team needs shared configs and an audit trail.
Get the occasional update
New adapters, releases, and the cloud waitlist. No spam, unsubscribe anytime.
Frequently asked questions
- What is mcpfold?
- mcpfold is a free, open-source CLI that keeps one canonical config for your MCP (Model Context Protocol) servers and folds it out to every client — Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Zed, and more — each in its own native format. You write your servers once; mcpfold renders the right file for each tool.
- How does mcpfold handle secrets?
- mcpfold stores secrets as references (for example ${env:GITHUB_PAT} or ${op:vault/item/field}), never as raw values. The reference is the only thing committed to git; the actual secret is resolved from your environment or secret manager at fold time, so credentials are never written to a client config on disk.
- Which MCP clients does mcpfold support?
- mcpfold folds to 18 clients from one config: Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Zed, Cline, Gemini CLI, JetBrains AI Assistant, Visual Studio, Continue, Roo Code, Goose, Codex CLI, LM Studio, Warp, opencode, and GitHub Copilot CLI. Each client has its own config format and path — including YAML (Goose) and TOML (Codex CLI) — and mcpfold handles the per-client dialect automatically.
- How does mcpfold reduce MCP token usage?
- mcpfold reduces MCP token usage by curating which tools each client loads, from one canonical config, so only the tools an agent actually needs enter its context window. Every connected MCP server otherwise dumps its full tool schema into context on every turn, used or not. In a reproducible benchmark, trimming a 45-tool setup to the 9 tools actually needed cut tool-schema tokens by about 80% (7,476 to 1,497), with no extra configuration.
- Do I still need mcpfold if my client has native tool-search?
- They complement each other. Native tool-search (such as Claude’s Tool Search Tool or OpenAI’s deferred tool loading) lets a model load tools on demand, but it is model-driven, non-deterministic, and only exists on the clients and models that ship it. mcpfold curates deterministically from one config across every MCP client — including Cursor, Windsurf, and Zed, which have no native tool-search — and trims the catalog before any native tool-search runs, so the two layers stack.
- Is mcpfold free?
- Yes. The entire mcpfold CLI — every adapter, the canonical config, secret references, sync, diff, doctor, and the config-as-code CI gate — is free and open source under the MIT license, with no account required. Optional paid cloud features add team config sharing, roles, and an audit trail.