Documentation

MCP Proxy

Executor is one MCP endpoint in front of all your integrations: connect once, and every agent gets the same catalog with shared auth and policies.

Executor sits between your agents and your tools as a single MCP endpoint. Your agents connect to Executor; Executor connects out to your integrations and re-exposes them as one catalog. Every tool call passes through the proxy, which is where auth and policy live.

Why proxy through Executor

  • One endpoint, every agent. Point Claude Code, Cursor, ChatGPT, or your own SDK at the same Executor endpoint instead of configuring each tool in each client.
  • Credentials stay out of the agent. A connection’s credentials are stored by Executor and attached to the upstream call. The agent runs in a sandbox and never sees them.
  • Per-tool policies. Every call is governed by a policy: allow, require approval, or block.
  • Mix integration types. Upstream MCP servers, OpenAPI specs, and GraphQL endpoints all show up in the same catalog.

How it works

Agents speak MCP to Executor. For each tool call, Executor picks the right integration, attaches that connection’s credentials to the upstream request, enforces the tool’s policy, and returns the result. The agent only ever talks to Executor, and the credentials never reach it.

Proxy an upstream MCP server

Add an MCP server as an integration and its tools join your catalog alongside everything else. Create a connection to it (with credentials if it needs them), and it becomes reachable through your one Executor endpoint, governed by the same policies as the rest of your tools.

This is the core of the proxy: your agents keep talking to a single endpoint while you add, swap, or remove upstream servers behind it, with no client-side change.

Connect your agents

Agents connect to Executor’s MCP endpoint. The exact command depends on how you run Executor:

Once a client is connected, every integration you add to Executor appears in that agent automatically.

Search and invoke mode

Add ?mode=passthrough to your MCP endpoint, or enable Search and invoke in the Connect card. This mode exposes four discovery and invocation tools:

  • integrations: list connected accounts with integration descriptions, account labels, and their last recorded health. Results are paginated, with one item per account. Health is not checked again by this call.
  • skills: list this server’s guides, or read skills({ name: "search-invoke" }).
  • search: find tools and read their JSON input schemas. Use the exact integration, owner, and connection fields from integrations to select an account.
  • invoke: call the exact tool ID from search with JSON arguments.

For example, call integrations({}), then search({ query: "create issue", integration: "github", owner: "org", connection: "main" }) using the account names returned by your server. Call invoke with the matching result’s id and arguments that follow its inputSchema.

Both integrations and search return items, total, hasMore, and nextOffset. To get another page, repeat the same query and filters with offset: nextOffset. This pagination finds tools and accounts; each upstream API has its own pagination for records.

The tool list stays small as you add integrations. Input schemas are loaded only for matching search results. Your client handles approval for invoke, and workspace block policies still apply. This mode does not expose a general code execution tool. Artifacts remain available unless you disable them with artifacts=false or the Artifacts toggle. Use skills({ name: "create-artifact" }) for the guide to building artifacts after discovering data with search and invoke.