AI library

Source · repo

MCPtoCodeMode: all your MCP servers behind three tools

Author
minkaworks
Published
22 October 2026
Added
23 October 2026
Original
https://github.com/minkaworks/mcptocodemode

Summary

MCPtoCodeMode (MIT) is a single MCP server that sits in front of the MCP servers you already use. Your agent sees three tools instead of every tool definition. To chain calls, the agent writes one short JavaScript program instead of making many separate tool calls. It appeared in today's GitHub trends (7 stars, first seen 2026-10-22; the publication date is approximate). It is young but well documented, with 28 commits, a CHANGELOG, a SPEC, a SECURITY policy, and CI.

Key points

  • The three tools: search finds methods by keyword and returns TypeScript signatures. describe gives the full signature of a server or method. execute runs JavaScript in which each server is a global, for example await notes.get({ id: "1" }). Only the return value, the console output, and the list of methods called go back to the model.
  • Sandbox: the code runs in a QuickJS WebAssembly isolate with no network, no filesystem, and no credentials. Its only way out is through the configured MCP servers. Each run is limited to 30 seconds and 100 tool calls.
  • Install:
    • Guided: npx -y mcptocodemode setup finds your installed agents and offers to edit their configs.
    • Claude Code: claude mcp add codemode -- npx -y mcptocodemode serve.
    • Codex: add [mcp_servers.codemode] with command = "npx" and args = ["-y", "mcptocodemode", "serve"] to ~/.codex/config.toml.
    • Any client: npx -y mcptocodemode serve --config /path/to/servers.json.
  • Tested clients (per the README): Claude Code 2.1, Codex 0.163, OpenCode 2.1, Pi 1.3, and Grok 1.0. Cursor's config format is covered by tests, but Cursor itself was not run.
  • Savings claim: about 5.7× fewer input tokens on the author's test tasks. This is self-reported, and the caveats are in REPORT.md.

Why it matters

It is a ready-made way to try the "code mode" pattern with many MCP servers: load tools on demand, and chain calls in code so intermediate results stay out of the context. Clients with built-in tool search (Claude Code has it on by default) already get part of this benefit. The main gain is for multi-step chains and for clients without tool search.