[IBJunior]

Every enabled MCP server used to be connected on each request, with all of its tools sent to the model. Context grew with every server, tool choice got worse, and each server cost a connection whether it was used or not.

With this change the model sees a short list of capabilities and one tool, load_tools. It searches a cached catalog of every server's tools and adds the best matches to the model's tools for the rest of the thread. A server connects only when one of its tools is called, through a pooled connection that closes after 5 idle minutes.

Full design: docs/MCP_DISCLOSURE.md.

What's in it

• packages/mcp-disclosure (@agentailor/mcp-disclosure), a workspace package with the catalog, BM25 search, connection pool and middleware. It imports no app code and no MCP client library; src/lib/mcp/connector.ts is the only agent-path file that uses @langchain/mcp-adapters.
• load_tools({ query, server?, limit? }) loads 5 matches by default, at most 10. An exact tool name loads only that tool, and a search with no match returns the tool names to try instead.
• Loaded tools are appended and never removed, because changing the tool list costs a prompt-cache miss. Tool schemas are sent to the model exactly as the server lists them.
• Servers that need sign-in are detected from stored OAuth state without connecting. load_tools never loads their tools and tells the model to ask for a reconnect on the Connections page.
• Server descriptions: an optional owner-written description on the Connections page, otherwise a one-line summary generated on the owner's model when the tool list changes.
• Tracing: chat spans record the tool definitions offered (with content capture on) and the provider's cache reads and writes.
• Migrations: 0004 adds the mcp_server_catalog table and 0005 adds mcp_server.description. Run pnpm db:migrate.

With no MCP server enabled, the model sees no load_tools and no capability list, so prompts and tools are unchanged.

Testing

• pnpm test, tsc --noEmit, pnpm typecheck:eval and pnpm build pass.
• Evals: 4 new cases in eval/cases/mcp.cases.mts, on in-process mock MCP servers (opt in with mcp: true). All 4 passed with one run each: searching before using a tool, finding a tool from different words, loading nothing for a ledger question, and reporting a server that needs sign-in. Other cases see no MCP servers.
• Tried by hand against a real Notion MCP server on the dev stack.

Known limits

• On a large catalog, BM25 doesn't always rank the right tool in the top 5. The model can search again with other words, but there is no eval yet for how often it has to.
• MCP tools still run without approval, as before this change. Gating them is planned separately.

🤖 Generated with Claude Code