🦞 ClawMetry

See your agent think. Real-time observability for 14 AI agent runtimes: OpenClaw, NVIDIA NemoClaw, Claude Code, OpenAI Codex & 10 more. One dashboard for your whole agent fleet.

🌐 Read this in: English · 简体中文 · 日本語 · 한국어 · Español · Português (BR) · Français · Deutsch · हिन्दी · العربية · Русский · more →

One command. Zero config. Auto-detects everything.

pip install clawmetry && clawmetry

Opens at http://localhost:8900 and you're done.

Works with 14 agent runtimes

ClawMetry started as observability for OpenClaw, and now meters your whole agent fleet in one dashboard, auto-detecting each runtime on your machine:

🦞 OpenClaw · 🟩 NVIDIA NemoClaw · ◆ Claude Code · ⬡ OpenAI Codex · Cursor · 🪿 Goose · ⚡ Hermes · opencode · ◈ Qwen Code · Aider · NanoClaw · PicoClaw · Pi · Deep Agents

OpenClaw and NemoClaw are free in the open-source app; the other runtimes light up with ClawMetry Cloud or a self-hosted Pro license. Switch runtimes from the header and every tab — cost, tokens, tools, traces — re-scopes to that runtime.

What You Get

• Flow — Live animated diagram showing messages flowing through channels, brain, tools, and back
• Overview — Health checks, activity heatmap, session counts, model info
• Usage — Token and cost tracking with daily/weekly/monthly breakdowns
• Sessions — Active agent sessions with model, tokens, last activity
• Crons — Scheduled jobs with status, next run, duration
• Logs — Color-coded real-time log streaming
• Memory — Browse SOUL.md, MEMORY.md, AGENTS.md, daily notes
• Transcripts — Chat-bubble UI for reading session histories
• Alerts — Budget caps, error-rate triggers, agent-offline detection; routes to Slack, Discord, PagerDuty, Telegram, Email
• Approvals — Gate destructive deletes, force pushes, DB mutations, sudo, package installs, network calls behind one-click sign-off

Screenshots

🧠 Brain — Live agent event stream

📊 Overview — Token usage & session summary

⚡ Flow — Real-time tool call feed

💰 Tokens — Cost breakdown by model & session

🧬 Memory — Workspace file browser

🔐 Security — Posture & audit log

🚨 Alerts — Budget caps, error-rate triggers, webhooks to Slack / Discord / PagerDuty / Email

✋ Approvals — Gate risky tool calls behind manual sign-off; policy-backed protection rules

Install

One-liner (recommended):

curl -sSL https://raw.githubusercontent.com/vivekchand/clawmetry/main/install.sh | bash

pip:

pip install clawmetry
clawmetry

From source:

git clone https://github.com/vivekchand/clawmetry.git
cd clawmetry && pip install flask && python3 dashboard.py

v2 Frontend Development

The v2 React app lives in frontend/ and is served at /v2 when the Flask server is started with v2 enabled.

Use two terminals while developing:

# Terminal 1: Flask API/server on :8900
CLAWMETRY_V2=1 python3 dashboard.py

# Terminal 2: Vite dev server on :5173
cd frontend
nvm use
npm ci
npm run dev

Open http://localhost:5173/v2/. Vite proxies /api requests to http://localhost:8900, so the React app can talk to the local Flask server without extra CORS setup.

To build the bundle that ships with the Python package:

cd frontend
npm run build

The production bundle is written to clawmetry/static/v2/dist/.

Runtime / Agent Compatibility

ClawMetry observes many AI-agent runtimes, not just OpenClaw. Each non-OpenClaw runtime ships a dedicated reader adapter that translates its native session format into ClawMetry's unified shapes; the daemon ingests them into the same DuckDB store + cloud snapshot, tagged with the runtime, and the Session replay tab shows a runtime switcher when more than one is present. See docs/compatibility.md for the full matrix + a guide to adding runtimes, and docs/RUNTIME_FAMILY.md for the OpenClaw-family primer.

"Beta adapter" means ClawMetry ships a reader for that runtime's real on-disk format, each built + verified against a real install on a real machine (see tests/fixtures/runtimes/<rt>/). Adapters are read-only; each is honest about what its runtime actually stores (e.g. PicoClaw/NanoClaw/Cursor don't write token cost to disk). When several runtimes run on one node, the runtime switcher scopes the sessions view to one for a clean deep-dive.

Track any SDK agent — out-loop cost attribution

The runtimes above all write sessions to disk. Your own production agent — the one you built on the OpenAI Agents SDK, LangChain, the Vercel AI SDK, LlamaIndex, E2B, or a plain httpx loop — doesn't. ClawMetry's zero-config interceptor still captures its LLM calls (cost, tokens, latency, errors) by monkey-patching httpx/requests:

import clawmetry.track            # activate the interceptor
clawmetry.track.set_source("support-agent")   # name this product

# ...your agent runs as normal; every LLM call is now tracked + attributed.

set_source() (or the CLAWMETRY_SOURCE=support-agent env var) tags each call with a named source, so every product you run shows up as its own first-class, cost-attributable line in the dashboard's 🔌 Out-loop sources card on Overview — calls, providers, latency, error rate per agent. No source set? The calls are still tracked; the card just stays hidden.

CLAWMETRY_SOURCE=billing-agent python my_agent.py

This is the same data layer the runtime adapters feed (DuckDB → cloud snapshot), so out-loop sources sync to the cloud dashboard the same as everything else, E2E-encrypted.

OpenTelemetry — vendor-neutral, send your traces anywhere

ClawMetry speaks OpenTelemetry in both directions, using the GenAI semantic conventions, so your agent traces are never locked into one tool.

Export every session — LLM calls, tools, sub-agents, tokens, cost — as OTLP/HTTP GenAI spans to any collector (Datadog, Grafana, Honeycomb, or your own OTel Collector):

clawmetry --otel-export http://localhost:4318/v1/traces
# equivalently:
CLAWMETRY_OTEL_EXPORT_ENDPOINT=http://localhost:4318/v1/traces