How do I connect DecisionPen MCP to Cursor or Claude?

What you are connecting

DecisionPen is the Architecture Context Plane. The Context Engine (MCP) lets Claude, Cursor, and other agents read that plane and publish evidence for human review. This guide is the setup path: create an Admin token, run the MCP server, then wire Cursor or Claude Desktop.

Full technical mirror on GitHub: CONNECT.md · MCP README.

Step 1 — Create an Edge API token (Admin)

You need a DecisionPen credential the MCP server can present to the API.

1. Sign in as an Administrator.
2. Open Admin → Security & audit.
3. Open Edge API tokens.
4. Create a token. Copy the secret once (dp_edge_…). You will not see it again.
5. Store it as DP_API_TOKEN in your edge environment. Never commit it to git.

For a quick laptop tryout you can paste a browser session JWT instead. Prefer Edge tokens for daily Cursor use, cron, or Docker. Revoke tokens anytime from the same Admin screen.

Step 2 — Install the MCP server

Use a dedicated Python virtualenv for MCP. Its dependencies conflict with the main DecisionPen app environment.

python -m venv .venv-mcp
# Windows
.\.venv-mcp\Scripts\Activate.ps1
pip install -r decisionpen_mcp\requirements.txt
copy decisionpen_mcp\.env.mcp.example .env.mcp

Minimum settings in .env.mcp:

DP_API_BASE=https://<your-decisionpen-host>
DP_API_TOKEN=dp_edge_...
DP_MCP_MODE=cite_only

Keep the DecisionPen API running (app venv). Start with cite_only so agents can read the plane before you enable publish.

Step 3 — Connect Cursor

Add a server to your Cursor MCP config. Paths must be absolute on Windows:

{
  "mcpServers": {
    "decisionpen": {
      "command": "C:\\path\\to\\Archimedes\\.venv-mcp\\Scripts\\python.exe",
      "args": ["-m", "decisionpen_mcp"],
      "env": {
        "PYTHONPATH": "C:\\path\\to\\Archimedes",
        "DP_API_BASE": "https://<your-decisionpen-host>",
        "DP_API_TOKEN": "dp_edge_...",
        "DP_MCP_MODE": "cite_only"
      }
    }
  }
}

Reload MCP in Cursor. Ask the agent to call describe_context_model, then try a landscape question with search_entities and get_entity_evidence.

Step 4 — Connect Claude Desktop

Use the same server shape in Claude Desktop MCP settings, then restart Claude.

{
  "mcpServers": {
    "decisionpen": {
      "command": "C:\\path\\to\\Archimedes\\.venv-mcp\\Scripts\\python.exe",
      "args": ["-m", "decisionpen_mcp"],
      "env": {
        "PYTHONPATH": "C:\\path\\to\\Archimedes",
        "DP_API_BASE": "https://<your-decisionpen-host>",
        "DP_API_TOKEN": "dp_edge_...",
        "DP_MCP_MODE": "cite_only"
      }
    }
  }
}

Step 5 — Enable publish (optional)

When agents should push ADRs, requirements, or notes into DecisionPen:

1. Set DP_MCP_MODE=propose and reload MCP.
2. Use submit_document (or S3 propose_enrichment).
3. Review proposals in the DecisionPen Inbox. Humans approve; the Confirmed map does not auto-update.

Step 6 — Optional: document repositories

To let agents list, search, and fetch under an allowlisted S3 prefix, see Connect an S3 document store via MCP and the GitHub runbook SYNC.md.

What comes next

Next in the Agents & MCP shelf: S3 document repositories, or continue the reading path with Beyond traditional EA.