Agent Control

MCP

MCP (Model Context Protocol) is a common "plug" standard for exposing tools in a consistent way. MCP servers are added to Tool Registry, and BaseLLM can itself expose a project's tools as an MCP server.

Connection

When you add an MCP server, its tools are listed automatically and enabled with admin approval. A tool that hasn't been approved is not visible to the model. Instead of a separate integration for every system that speaks MCP, you use one uniform connection.

  • Name, description, input and output schema
  • Risk level and allowed projects
  • Human approval setting
  • Per-minute call limit and timeout
  • Audit Log setting and policy rules

Using BaseLLM as an MCP server

An MCP client connects to a project's granted tools at POST /api/v1/mcp, sending an Authorization: Bearer <project API key> header. The server is stateless and uses Streamable HTTP; only the tools granted to the project are listed, and the same approval and logging rules apply.

Security principles

  • Your organization adds every tool; no tools are downloaded at random from the internet
  • Every call goes through BaseLLM: permissions, approval, masking
  • Tool output is labeled as data; it can't change permissions
  • Connection secrets are stored with AES-256-GCM under a separate key
  • Audit Log records events, not content