Executor Collapses 278,800 Tokens of AI Tools Into 1,044 With One Gateway
A new open-source MCP gateway lets one catalog of tools serve every agent, cutting a 278k-token prompt down to about 1k tokens.
PRO- Executor is an open-source MCP gateway that unifies OpenAPI, GraphQL, and MCP tools into one catalog.
- Presents a single code-mode execute tool, cutting 1,640 tool definitions from ~278k tokens to ~1k.
- Sandboxed JS runtime injects credentials host-side so agents and models never see raw tokens.
- Per-tool policies: allow, require approval, or block, with defaults derived from the source spec.
- Runs as Cloud, CLI, desktop app, Docker, or Cloudflare Worker; MIT licensed and YC-backed.
- Free tier covers 10k executions/month; Team plan is $150/org/month for 250k executions.
Anyone running more than one coding agent has felt the friction: Claude Code, Cursor, ChatGPT, and Codex each need their own copy of every integration, the same API keys pasted three times over, and no shared notion of which tools are safe to run unattended. Executor is an open-source project trying to collapse that mess into a single catalog, and it has picked up roughly 3.1k GitHub stars along the way.
Configure every integration once (MCP servers, OpenAPI specs, GraphQL APIs) with authentication and per-tool policies, then use that one catalog from any MCP-compatible agent. Everything downstream, from Claude Code to a self-written CLI script, talks to the same gateway.
The token math that makes it interesting
The most striking technical claim is about context. Wire up a realistic enterprise stack and you drown the model in tool definitions. Without Executor, 1,640 tools across GitHub, Stripe, Jira, Sentry, Linear, Gmail, Notion, and Slack cost roughly 278,800 tokens. With Executor, the agent sees a single tool costing about 1,044 tokens.
How? Executor exposes what it calls a code-mode MCP. Instead of enumerating every function up front, it gives the model one execute tool that runs sandboxed TypeScript, plus tools.search and tools.describe helpers. The model searches the catalog and loads a tool's schema only when the code actually calls it, so the prompt never balloons. A typical workflow inside the sandbox looks like this:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
