AgentHub Monitor

An MCP App server: it renders an interactive pipeline dashboard inside the host window (Claude Desktop, Claude Code, and other MCP App–capable clients) instead of returning a wall of text.

The dashboard shows two things side by side:

DAG state — nodes of a pipeline run, grouped into waves, each with its own status.
Harness metrics — scenario pass rate, critical failures, accumulated cost and latency.

Point it at your own pipeline's JSON export and it becomes a live view of your runs. With no configuration it serves a built-in sample, so the UI is useful the moment you install it.

Tools

Tool	What it does	Read-only
`show_agenthub_monitor`	Renders the dashboard inline in the host	yes
`get_status`	Returns the current pipeline state as structured JSON	yes
`trigger_action`	Records an action request in the server's own state	no

trigger_action does not execute anything. It annotates the server's in-memory state so the dashboard can show that an action was requested. This server runs no commands, spawns no processes, and makes no network calls.

Install

Requires Python 3.10+. No third-party dependencies.

git clone https://github.com/NEVELSK65/mcp-agenthub-monitor.git
cd mcp-agenthub-monitor
python3 -m unittest discover -s tests    # optional: verify

Configure

Add to your MCP client configuration:

{
  "mcpServers": {
    "agenthub-monitor": {
      "command": "python3",
      "args": ["/absolute/path/to/mcp-agenthub-monitor/server.py"],
      "env": {
        "AGENTHUB_STATE_FILE": "/absolute/path/to/your/pipeline-state.json"
      }
    }
  }
}

AGENTHUB_STATE_FILE is optional. Without it the server returns sample data, flagged with is_sample_data: true so it is never mistaken for a real run.

State file format

Write this JSON from your own CI, orchestrator, or test harness:

{
  "dag": {
    "title": "Nightly build",
    "total_nodes": 6,
    "completed_nodes": 4,
    "waves": 3,
    "node_states": {
      "node_01_checkout": "COMPLETED",
      "node_02_unit_tests": "COMPLETED",
      "node_03_integration": "RUNNING",
      "node_04_security_scan": "PENDING"
    }
  },
  "harness": {
    "pass_rate_pct": 93.75,
    "total_scenarios": 16,
    "passed_scenarios": 15,
    "critical_failures": 0,
    "total_cost_usd": 0.112,
    "total_latency_ms": 41700
  },
  "log": "Integration stage in progress."
}

The file is re-read on every tool call, so the dashboard reflects whatever your pipeline last wrote. Node status strings are passed through to the UI as-is; COMPLETED, RUNNING, PENDING and FAILED get distinct colours.

Resource

The UI is exposed as an MCP resource at ui://agenthub/mcp-app.html with MIME type text/html;profile=mcp-app. Hosts that support MCP Apps render it inline; hosts that do not can still read it as an HTML document.

Privacy

The server reads one local file (the one you name in AGENTHUB_STATE_FILE) and sends nothing anywhere. See PRIVACY.md.

License

MIT — see LICENSE.