cueapi-mcp

Official Model Context Protocol server for CueAPI, the open-source coordination layer for AI agent systems.

Give your MCP-enabled assistant (Claude Desktop, Cursor, Zed, or any other MCP host) the ability to schedule agent work, fetch execution history, and close the loop with evidence-backed outcome reports, all from inside a conversation.

Why

Agents don't finish in one call. They coordinate across time, tools, environments, agents, and humans. Every handoff is a place where silent failure hides. CueAPI closes each handoff with structured evidence: an external ID, a result URL, or an artifact. This MCP server gives the agent direct access to that surface, so the agent can both schedule its own follow-up work and report outcomes with proof.

Install

npm install -g @cueapi/mcp
# or use via npx (no install):
npx -y @cueapi/mcp

Configure (Claude Desktop)

Add this to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "cueapi": {
      "command": "npx",
      "args": ["-y", "@cueapi/mcp"],
      "env": {
        "CUEAPI_API_KEY": "cue_sk_..."
      }
    }
  }
}

Generate your API key at cueapi.ai. Self-hosting? Set CUEAPI_BASE_URL alongside CUEAPI_API_KEY.

Configure (Cursor / Zed / other hosts)

Any MCP host that supports stdio servers can run this. Point the host at the cueapi-mcp binary and pass CUEAPI_API_KEY in the environment.

Tools exposed

Example conversation

You: Schedule a daily 9am job that posts a digest to my webhook.

Assistant (uses cueapi_create_cue): Created cue cue_abc123, first fire tomorrow at 9:00 UTC.

You: Show me the last five times it ran.

Assistant (uses cueapi_list_executions): ...

Development

npm install
npm test        # vitest smoke tests for the tool surface
npm run build   # compile TypeScript to dist/
npm run dev     # run the server locally with tsx

Links

• CueAPI homepage: https://cueapi.ai
• Docs: https://docs.cueapi.ai
• Core (open source): https://github.com/cueapi/cueapi-core
• Model Context Protocol: https://modelcontextprotocol.io

Changelog

• 0.4.1. Add cueapi_update_cue tool: update a cue's mutable fields (name, schedule, callback URL, timezone, payload, description). Wraps PATCH /v1/cues/{id}. Sparse update — only provided fields are sent. Closes the cross-wrapper parity gap (every other client wrapper exposed update; MCP didn't until now).
• 0.4.0. Add five execution-lifecycle tools — cueapi_get_execution, cueapi_list_claimable_executions, cueapi_claim_execution, cueapi_claim_next_execution, cueapi_execution_heartbeat. Closes the receive-claim-process-complete loop for MCP-host agents that want to consume worker-transport executions from in-session (e.g. Claude Desktop, Cursor, Zed). Highlights: list_claimable_executions filters server-side via task / agent query params (client-side filtering hits a known LIMIT-50 starvation bug); claim_next_execution accepts an optional task_name and internally fans out (filtered list → pick oldest → claim by ID) since the server doesn't yet support a task filter on the bare claim endpoint; execution_heartbeat sends worker_id via the X-Worker-Id request header (the server's transport for that field) and requires it in the schema so misconfigured callers fail at the wrapper instead of silently bypassing race protection. Internal: CueAPIClient.request() gains an optional extraHeaders parameter to support per-call custom headers.
• 0.3.0. Add cueapi_fire_cue tool: fire an existing cue immediately with an optional payload_override (and merge_strategy: 'merge' | 'replace', default 'merge'). Wraps POST /v1/cues/{id}/fire. Lets agents trigger ad-hoc one-shot executions without creating throwaway cues, and lets per-fire dynamic data flow through to webhook dispatch + worker-claim responses without mutating the stored cue.
• 0.1.4. Fix cueapi_pause_cue / cueapi_resume_cue to use PATCH /v1/cues/{id} with {"status": "paused" | "active"} (previously called non-existent /pause and /resume endpoints, returning a runtime 404). PR #1. This is the release that actually contains the fix; 0.1.3 was published prematurely with this note but without the merged code.
• 0.1.3. Premature publish, superseded by 0.1.4. No functional changes from 0.1.2.
• 0.1.2. Register with the Official MCP Registry.
• 0.1.0. Initial release: 8 tools for create / list / get / pause / resume / delete cues, list executions, report outcome.

License

MIT © Vector Apps Inc.