← Back to architecture

MCP

Model Context Protocol · knowledge inside someone else's AI

Company knowledge — inside Claude Code

Someone else's AI assistant queries your knowledge base on its own and weaves the answer into the work — without leaving the editor.

The client's model does the thinking, not ours. MCP hands it one tool — search_knowledge returns findings, not a finished answer; the assistant synthesizes on its own. Search runs under the owner's identity →

01

Setup

The server is networked — the client just remembers an address and a key. One command.

claude mcp add --transport http achilles https://achilles.local/mcp \
  --header "Authorization: Bearer ach_…"

With OAuth (v2) the key drops out — the command becomes a single step. MCP is a reference surface of Public API for AI clients; the /mcp path is on the shared HTTP API map.

02

Authorization

No login of its own — the client resolves to a user_id, then the owner's role and ACL.

  1. API key ach_… in the header · headless · read-only
  2. v2 OAuth 2.1 browser consent, no key to carry

ACL sits on top: the key's scope only narrows the rights.

03

Governing the door

Entry from outside in — control at two levels.

  • Platform kill switch. A single toggle in platform settings closes MCP for everyone at once — the door is locked, the keys stay alive.
  • Keys are self-service. An employee creates their own — in their profile; Owner/Admin oversight on the keys screen. Deactivation — the keys go dark.

Every call is rate-limited and tagged to a key — the keys screen shows usage; request text is not logged.