Cursor & OpenClaw

Cursor IDE

  1. Bring up the stack: ./scripts/openfdd_stack_up.sh standalone (or Railway private central)
  2. Obtain an operator JWT: POST /api/auth/login with {"username":"agent","password":"$OPENFDD_AGENT_PASSWORD"}
    (or admin bearer → POST /api/auth/agent-token). Do not put OPENFDD_ADMIN_PASSWORD in MCP config.
  3. Configure the MCP server (ghcr.io/bbartling/openfdd-mcp) with OPENFDD_API_BASE → central and OPENFDD_MCP_TOKEN → that JWT
  4. Use tools for health, assignments, CSV preflight, rules, reports

Railway: see RAILWAY_DEPLOYMENT.md § Secure agent auth — secrets stay in Railway Variables; MCP stays private.

Open-FDD does not ship an in-dashboard chatbot. Connect Cursor to openfdd-mcp (stdio) — see external agents examples.

OpenClaw on Raspberry Pi

OpenClaw agents on a fresh Pi bring up the stack from GHCR images — see Raspberry Pi edge:

ghcr.io/bbartling/openfdd-central:latest
ghcr.io/bbartling/openfdd-web
ghcr.io/bbartling/openfdd-fieldbus:latest
ghcr.io/bbartling/openfdd-mqtt:latest

Typical flow:

  1. ./scripts/openfdd_stack_up.sh standalone on the Pi
  2. Validate with ./scripts/openfdd_health_check.sh
  3. Wire MCP with agent (operator) JWT, not admin password
  4. Use agent tools for commissioning — never BACnet writes without approval

Grounded workflows

Task MCP / API
CSV import openfdd_csv_import_preflight, /api/csv/import/execute
Hourly CSV append openfdd_csv_package_append → POST /api/csv/import/package/append (JWT + confirm:true)
Model bind /api/model/assignments/save
Rule deploy /api/rules/batch, /api/fdd-rules/{id}/activate
Site health /api/health, /api/dashboard/summary

Avoid speculative language — verify with health and historian plots before claiming faults are active.