Skip to main content

​
What it is

The Knowledge Base MCP server exposes semantic search and retrieval over the Knowledge Base as a Model Context Protocol server. It is a separate, independently deployed service from the Framework MCP server — same connection method, same sign-in, but a different service with a narrower purpose.

​
Why two MCP servers

The Framework MCP server manages planning: projects, checklist items, artifacts, users, and organizations. The Knowledge Base MCP server does one thing — semantic search and retrieval of Knowledge Base Documents (decisions, guides, pitfalls, process docs). They’re split into separate services deliberately:
  • Independent deployment and database. The Knowledge Base has its own database and is versioned and deployed independently of the Framework. Splitting the MCP surface the same way keeps that boundary intact end to end, rather than giving the Framework server direct database access into a service it doesn’t own.
  • Different clients, different needs. An agent that only needs to look up prior art doesn’t need — and shouldn’t be granted — the full planning and administration toolset (managing users, organizations, departments, and projects).
In practice, you rarely need to choose one or the other in isolation:
  • If you’re already using the Framework MCP server for planning (writing ADRs, reviewing artifacts, checking planning status), you don’t need a separate connection at all — the Framework MCP server’s own knowledge_ask tool proxies into the Knowledge Base MCP server for you, forwarding your identity so results are already filtered to what you’re authorized to see. The same underlying connection is also what surfaces related prior art when the assistant looks up context for a checklist item.
  • If you want an agent to search or retrieve Knowledge Base content directly — outside of a planning session, or from a tool/client that doesn’t otherwise talk to the Framework MCP server — connect to the Knowledge Base MCP server directly, as described in Knowledge Base MCP Server Setup.

​
Authentication

Like the Framework MCP server, every request requires a valid sign-in — the same account and organization you use elsewhere in Engineering Framework, just authenticating against a different service.

​
Authorization

Every result is checked against the same permission model used by the dashboard’s search and review pages. An MCP caller only ever sees Documents that person is actually authorized to view — unauthorized results are simply left out of the response, never returned in a partial or redacted form.

​
Tools

The Knowledge Base MCP server registers three tools:
  • ask — semantic search over the Knowledge Base.
  • get_by_id — full Document retrieval by ID.
  • kb_health_check — a diagnostic/scaffold tool confirming the connection and auth are working; not a real Knowledge Base capability.
For parameters and behavior, see Searching & Asking the Knowledge Base, which documents ask and get_by_id in full — this page won’t repeat that detail.

​
Getting connected

See Knowledge Base MCP Server Setup for the Claude Code connection steps. Other MCP clients connect the same way, over the /mcp StreamableHTTP endpoint with a bearer token, using their own MCP server configuration UI. Looking to add content rather than search it? See the engframe CLI — a separate, standalone command-line tool for ingesting Documents.