Skip to main content

​
What it is

The Framework MCP server exposes Engineering Framework’s planning and administration operations as a Model Context Protocol server. It lets any MCP-compatible AI client — Claude Code, Claude Desktop, or another agent — read and write the same data your team manages through the web app, using the same authentication and authorization rules. There is no separate code path or relaxed permission model for agent access: every tool call is authenticated as a real user and goes through the exact same permission checks and business logic the web dashboard uses.

​
Why connect an agent to it

Engineering Framework organizes project planning around a checklist-style framework: categories of concerns (technical design, reliability, security, and so on) broken into individual checklist items, each of which can have a definition, an artifact template, and a completion status per project. Working through that framework by hand — figuring out what’s missing, writing the artifact, updating the record — is exactly the kind of multi-step, context-heavy task an AI agent is good at. Connecting an agent to the MCP server lets you, from your terminal or chat client:
  • Create a project and scope which framework items apply to it
  • Ask “what’s missing for my current phase?” and get a prioritized list of gaps
  • Generate a planning artifact (an ADR, rollback plan, SLO doc, etc.) from the project’s context and the item’s template
  • Review an existing artifact against its definition and template for gaps
  • Manage users, departments, and roles without leaving your agent session
The server pairs tools (individual operations) with prompts (multi-step guided workflows that chain several tool calls together with instructions for the agent to follow).

​
How it fits into the product

The MCP server and the web dashboard are two doors into the same house: an AI agent connects to the MCP server, which enforces the same authentication and permissions as the dashboard, and reads and writes the same underlying data.
  • Authentication: the agent’s MCP client presents a valid bearer token on every request, the same kind of token a browser session uses.
  • Authorization: each tool call is checked against the same permission model a dashboard user would go through. A tool is hidden entirely from an agent that can’t view it, and rejected at execution time if the caller can’t actually perform it — an agent never sees or successfully calls a tool outside that person’s real permissions.
  • Data: all reads and writes go through the same underlying logic the web app uses, so state stays consistent regardless of whether a person or an agent made the change.

​
Getting connected

See Claude Code Setup for the specific steps to add the Framework MCP server to Claude Code, including the CLI command, JSON config, and a CLAUDE.md snippet your organization settings page can generate for you pre-filled with your organization (and optionally project) IDs. Other MCP clients (Claude Desktop, etc.) connect the same way, over the /mcp StreamableHTTP endpoint, using their own MCP server configuration UI. Once connected, see:
  • Tools Reference for every individual tool, its parameters, and what it returns
  • Prompts Reference for the five guided workflows and the natural-language phrasing that triggers each one
There’s a second, separate MCP server for the Knowledge Base — search and retrieval over Knowledge Base Documents, not planning or administration. The Framework MCP server’s own knowledge_ask tool already proxies into it, so you often don’t need a separate connection. See Knowledge Base MCP Server: Overview for when you’d connect to it directly instead.