MCP Server Overview

What is the Closedloop MCP Server?

The Closedloop MCP server exposes the Closedloop platform to AI agents through the Model Context Protocol. Any MCP-compatible agent — Claude Code, Claude Desktop, Cursor, or your own toolchain — can connect to the server and interact with projects, documents, loops, and more using natural language.

The server is hosted at https://mcp.closedloop.ai/mcp and authenticates via OAuth using your Closedloop API key.

Connecting an AI Agent

The steps below apply to Claude Code CLI. Other MCP clients follow the same pattern: provide the server URL and complete the OAuth flow.

1. Run /mcp in the Claude Code CLI.
2. Add a new server:
   • Name: closedloop
   • URL: https://mcp.closedloop.ai/mcp
3. Choose Authenticate (or Re-authenticate if reconnecting).
4. Complete the browser OAuth flow and enter your Closedloop API key (sk_live_...).
5. Confirm the /mcp panel shows Status: connected and Auth: authenticated.
6. Test with a prompt such as: list my projects.

Entity Hierarchy

All work in Closedloop is organized as projects containing documents:

Project (PRO-*)
  └── Document (PRD-* | PLN-* | FEA-*)

Tools that accept an entity reference accept both the UUID and the human-readable slug. Pass the slug verbatim — the API resolves it server-side.

Available Tools

The MCP server provides tools organized into the following categories. It also registers a ping connectivity check used when verifying a new connection.

Document Management

Artifact Links

Loops

Projects

Tags

Search

Comment Threads

User and Status

Audit Ledger

verify-audit-ledger takes no input — the organization is resolved from the API key you authenticated with, so a caller can only ever verify its own ledger. It returns JSON:

{
  "ok": true,
  "head": {
    "seq": "42",
    "hash": "9f86d081884c7d659a2feaa0c55ad015a3bf4f1b2b0b822cd15d6c15b0f00a08"
  }
}

head is the highest entry in the chain: seq is a decimal string (the sequence number is a 64-bit integer that does not survive JSON) and hash is its 64-character hex chain hash. An organization with no entries yet returns the empty sentinel — seq "0" and a hash of 64 zeros.

When verification fails the tool returns ok: false alongside brokenAtSeq — the seq of the first row that fails a check, which is not necessarily a row whose own hash fails to recompute — and a machine-readable reason, one of:

The three run in that order per row, so a sequence_gap or broken_link is reported against a row whose own contents may recompute perfectly — the break is in the chain, not in that row. head is still included, read separately so you can see how far the chain has advanced past the break.

Agent Sessions

Agent Components

Status Lifecycles

Documents and Features have separate status vocabularies — use the one that matches the artifact type. The server rejects a status that doesn't belong to the target artifact's vocabulary.

Documents (PRD, Plan) — an authoring / approval lifecycle:

DRAFT → IN_REVIEW → APPROVED

APPROVED is sign-off (terminal). Use CHANGES_REQUESTED when a review asks for rework (return to IN_REVIEW after), EXECUTED for Implementation Plans that have been executed, and OBSOLETE for deprecated documents.

Features (FEA) — a delivery lifecycle:

TRIAGE → BACKLOG → TODO → IN_PROGRESS → IN_REVIEW → DONE

Update status as work progresses. For Features, agents should set IN_PROGRESS when starting work, IN_REVIEW when a PR is created, and DONE when merged; use BLOCKED when stuck and CANCELED for won't-do.

Next Steps