The Harness Watch MCP server lets an outside agent, such as Claude Code, Cursor or a reviewer bot, follow a harness without being able to change it. Every tool on it only reads.
Connect
You need an API key from your workspace.
| Setting | Value |
|---|---|
| URL | https://<your autobotAI API host>/mcp/harness-watch |
| Transport | HTTP (streamable) |
| Header | Authorization: ApiKey <key id>:<key secret> |
For Claude Code, add it to .mcp.json:
json{ "mcpServers": { "autobotai-harness-watch": { "type": "http", "url": "https://<your autobotAI API host>/mcp/harness-watch", "headers": { "Authorization": "ApiKey <key id>:<key secret>" } } } }
Connect a watching agent to this server only. Other autobotAI MCP servers include tools that change bots, harnesses and tasks.
Tools
| Tool | What it returns |
|---|---|
watch_list_harnesses | Every harness with its state, version and problems |
watch_harness_status | Schedules, each stage with its bot and task counts, findings waiting for a group, stuck tasks and health |
watch_harness_tasks | The harness's tasks with stage, status, group, linked pull request and deadline |
watch_task | One task in full: outputs per stage, journey with every routing decision, timeline, group members and what its runs did |
watch_harness_runs | Every run of the harness's bots, including runs started outside the harness, with who started it, state, duration and errors |
watch_run | One run: state, errors, step statuses and the last messages of the bot's conversation |
watch_harness_activity | One feed of everything since a time: task moves, runs started and finished, and actions taken. Pass the returned next_since back to get only what's new. |
Ask your agent
Once connected, ask in plain words, for example:
- "Check the payments-api harness: is anything stuck or failing?"
- "What happened in the harness in the last six hours?"
- "Why did task TASK-1042 go to escalation? Read its journey and the fix run's conversation."
- "Did any bot of this harness run outside the harness today?"