OpenClaw Observer for Claude CoWork

OpenClaw Observer: Production Observability Dashboard

Generate an interactive HTML dashboard that gives a complete picture of an OpenClaw deployment's health in one glance. The dashboard pulls data from three openclaw CLI commands and renders it as a visual, clickable report.

Prerequisites

This skill requires Desktop Commander (or equivalent terminal access) to run openclaw CLI commands on the user's machine. The dashboard script requires Python 3.8+ with only standard library modules — no pip installs needed.

How It Works

Step 1: Collect Data

Run these three commands via Desktop Commander and save the JSON output:

# Gateway health + channel status
openclaw health --json > /tmp/oc-health.json

# Deep status with session info
openclaw status --json --deep > /tmp/oc-status.json

# All sessions across all agents
openclaw sessions --json --all-agents > /tmp/oc-sessions.json

If any command fails (e.g., gateway not running), note the failure — the dashboard handles partial data gracefully.

Step 2: Generate the Dashboard

Run the bundled Python script:

python3 ${CLAUDE_SKILL_DIR}/scripts/generate_dashboard.py \
  --health /tmp/oc-health.json \
  --status /tmp/oc-status.json \
  --sessions /tmp/oc-sessions.json \
  --output /tmp/openclaw-dashboard.html

The script produces a self-contained HTML file (no external dependencies) with an interactive dashboard.

Step 3: Present to the User

Open the dashboard for the user. In Cowork, save it to the workspace folder and provide a computer:// link. In Claude Code, use open /tmp/openclaw-dashboard.html.

Then give a brief verbal summary highlighting:

• Any channels that are configured but not running (potential issues)
• Agents with high token usage (cost concerns)
• Sessions that were aborted or errored
• Whether heartbeats are enabled and firing

What the Dashboard Shows

Header Bar

• OpenClaw version, gateway uptime, and timestamp
• Overall health status (green/yellow/red)

Channel Health Panel

• Status badges for each configured channel (Telegram, Discord, iMessage, etc.)
• Running/stopped state, last error, and last probe time
• Color-coded: green = running, yellow = configured but stopped, red = error

Agent Overview Panel

• Card for each agent showing: heartbeat status, session count, total tokens used
• Sorted by activity (most active first)
• Heartbeat indicators: green = enabled + firing, gray = disabled

Session Activity Panel

• Table of recent sessions sorted by last activity
• Shows: agent, channel, session age, model, token counts, thinking level
• Highlights aborted sessions in red
• Filterable by agent

Token Usage Summary

• Bar chart showing token consumption by agent
• Total tokens across all sessions
• Model distribution (which models are being used and how much)

System Info

• Runtime version, default model, context window size
• Session store paths and counts per agent

Interpreting Results

After presenting the dashboard, proactively flag these patterns:

Red flags (action needed):

• Channels configured but not running — suggests gateway needs restart
• Sessions with abortedLastRun: true — the agent hit an error mid-response
• High token usage on a single session — possible context bloat, may need compaction
• Channel lastError is non-null — investigate the error

Yellow flags (worth monitoring):

• Agents with heartbeat disabled — they won't self-check
• Sessions older than 24 hours with no activity — may be stale
• Token count approaching context window limit — risk of degraded responses

Green signals (all good):

• All channels running with recent probes
• Heartbeats firing on schedule
• Token usage distributed evenly across agents

Scheduling Recurring Checks

If the user wants regular observability, suggest setting up a scheduled task:

/schedule openclaw-observer every 6 hours

Or guide them to add a cron in OpenClaw itself that generates the dashboard:

openclaw cron add --name "health-dashboard" --every "6h" --command "python3 <path>/generate_dashboard.py ..."