vivekchand/clawmetry

🦞 ClawMetry – Real‑time observability for AI agents

What it is – A locally‑run dashboard that automatically discovers the AI‑agent runtimes you have on a machine (e.g., OpenClaw, NVIDIA NemoClaw, Claude Code, OpenAI Codex, etc.) and shows, in real time, what each agent is doing, how much it costs, and whether anything looks wrong.

Why it matters – When you run many autonomous agents, you quickly lose visibility into:

• which LLM calls are being made,
• how many tokens and dollars each call consumes,
• whether a tool‑call is risky, and
• if an agent is stuck or blowing its context window. ClawMetry surfaces all of that in a single UI without requiring you to modify the agents themselves.

---

Key capabilities (as described in the README)

Feature	What you get
Automatic discovery	One‑command install (pip install clawmetry && clawmetry) finds any of the 30 supported runtimes on the host and reads their session files read‑only.
Sessions & transcripts	Turn‑by‑turn view of each agent’s conversation, with a replay button.
Cost & token tracking	Per‑runtime, per‑model, per‑day and lifetime spend, plus anomaly flags for spikes.
Live flow diagram	Visual graph of messages moving through channels, models and tools, updating as the agent runs.
Context‑window monitoring	Shows window utilisation, compaction events and “blowout” warnings, using a provider‑specific window table.
Memory & skill view	Lists files and skills each runtime has loaded.
Health & logs	Disk/memory usage, error rates, rate‑limit status, live log stream.
Alerts & approvals	Built‑in detectors (cost spikes, quiet agents, error bursts, etc.) and optional gating of risky tool calls that can be approved from a phone.
Export	OpenTelemetry (OTLP) support for sending traces to external back‑ends.
SDK tracking	If you build an agent on an SDK, the interceptor can attribute LLM calls to your own code.

---

Supported runtimes (free vs. paid)

• Free open‑source runtimes – OpenClaw, NVIDIA NemoClaw, Goose.
• Paid runtimes – Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Gemini CLI, Replit Agent, OpenHands, etc. (total > 30 agents).

---

Pricing & deployment

Plan	Coverage	Price (per node)
Free	Only the three open‑source runtimes, local‑only dashboard.	$0
Starter	All runtimes, fleet view, cloud sync.	$9 / month
Pro	Starter + approvals, policy engine, evaluation tools, anomaly detection, cost optimizer, OTel export, tamper‑evident audit log.	$19 / month

Self‑hosted license keys let you run the paid features without the SaaS component (clawmetry license).

---

Data privacy & egress

• By default ClawMetry reads local session files; no session data is sent off‑machine unless you run clawmetry connect.
• The optional connect flow encrypts data end‑to‑end in the browser; the encryption key never leaves the host.
• The only outbound traffic without opt‑out is an anonymous install ping, a PyPI version check, and a one‑time public‑IP lookup for a banner.
• All egress destinations are documented in docs/EGRESS.md and can be disabled.

---

Performance impact

• The daemon runs as a separate process; session‑file tailing adds 0 ms.
• Optional HTTP interceptor adds ≈0.44 ms per LLM call (≈0.009 % of a 5 s call).
• Optional pre‑tool hook adds ≈44 ms per gated tool call.
• Enforcement proxy adds ≈9.7 ms per LLM call.
• In a busy install the daemon consumes about 12 % of a single CPU core and ~70 MB per 100 k events.
• Overhead can be measured locally with python -m benchmarks.overhead.

---

Installation & quick start

pip install clawmetry      # installs the Python package
clawmetry                  # launches the dashboard at http://localhost:8900

• Works on Python 3.8+ on macOS, Linux or Windows.
• Docker image and compose file are provided (docs/DOCKER.md).

---

Documentation highlights

• Compatibility – what each adapter reads and how to add a new runtime.
• Context blowout – per‑provider window sizes, compaction vs. overflow.
• Overhead – detailed benchmark methodology.
• Approvals & policies – risk scoring and phone‑based approval flow.
• OpenTelemetry – exporting traces to any OTLP collector.
• SDK tracking – attributing costs for custom‑built agents.

---

License

MIT – free to use, modify, and redistribute.

---

TL;DR

ClawMetry is a self‑hosted observability platform that automatically discovers a wide range of AI‑agent runtimes on your machine and gives you a unified, real‑time view of their activity, costs, context usage, and health, with optional security‑focused approvals and cloud‑sync for a modest per‑node fee.