Y Combinator's Executor Shrinks 278,800 AI Agent Tokens Down to 1,044

A new open-source project called Executor targets one of the messiest parts of running AI agents at scale: every client (Claude Code, Cursor, ChatGPT, Codex) needs its own copy of every integration, its own API keys, and its own idea of what each tool is allowed to do. Executor ships as an MCP gateway that sits between your agents and your services, with a single catalog, shared credentials, and per-tool policies.

The repo has climbed past 3.1k stars and is backed by Y Combinator. It is MIT licensed, written in TypeScript on Bun and Turborepo, and can run as a hosted cloud service, a desktop app, a CLI daemon, a Docker container, or a Cloudflare Worker.

Token bloat, meet code mode

Wiring lots of MCP servers directly into a model destroys the context window. Executor's own benchmark claims that exposing roughly 1,640 tools from services like GitHub, Stripe, Jira, Sentry, Linear, Gmail, Notion, and Slack burns around 278,800 tokens, while routing through Executor exposes a single tool at about 1,044 tokens.

The mechanism is a code-mode MCP surface. Rather than dumping every tool schema into the prompt, the model sees one execute tool that runs TypeScript in a sandbox. The agent searches the catalog, fetches a tool's schema only when it needs it, then calls it:

const { items } = await tools.search({ query });
const path = items[0]?.path;
const details = await tools.describe.tool({ path });
const result = await tools[path](input);