MCP

Model Context Protocol (MCP) is how Jinn supplies typed tools to compatible agent engines. For each session, the resolver combines global configuration, engine capability, employee selection, and Jinn-specific overrides into a temporary MCP config.

The built-in jinn server exposes company primitives such as roster discovery, sessions, delegation, Todos, Workflows, cron reads, knowledge, costs, and managed files. It wraps the same gateway authority boundaries documented in the API; a typed tool does not grant operator powers to an employee session.

mcp:
  gateway:
    enabled: true
    engines:
      grok: false
  custom:
    internal-search:
      enabled: true
      command: node
      args: ["server.mjs"]
      env:
        SEARCH_TOKEN: ${SEARCH_TOKEN}

The gateway master switch, per-engine opt-out, authenticated smoke check, engine MCP capability, and employee overrides all participate. mcp.gateway.enabled: false is a kill switch. jinnMcp: true can pilot the company toolset for one employee when the global setting is absent, but it cannot override the global kill switch or an engine opt-out.

Employee mcp can be false, true, or a list of named servers. Custom servers may be stdio (command, args, env) or URL-based (url, optional headers).

Security and limits

Environment references are expanded into the session config; do not write literal secrets into public persona or skill files. The spawned built-in server receives a session-scoped capability and caller identity. If that identity is lost, sensitive mutation routes fail closed rather than becoming operator calls.

MCP availability is engine-dependent. A configured tool can still be absent when the engine lacks a per-session MCP mechanism or the startup auth smoke test fails. Check the session context and gateway logs instead of assuming config implies attachment.