ClawMetry
One command, zero config: real-time observability and governance for 30 AI agent runtimes — session replay, token costs, tool calls, anomaly alerts, and risky-call approvals.
Evidence shows read-only by default, 127.0.0.1 binding, local DuckDB, AES-256-GCM E2E encryption, approval gating off by default, and CI workflows with explicit least-privilege permissions and SHA-pinned actions — least_privilege earns full marks. Deducted for: egress documentation (EGRESS.md) is referenced but absent from reviewed files, so opt-out telemetry claims rest on README assertions; browser-decryption integrity is self-admittedly ineffective against the vendor; rollback capability is essentially unaddressed; the closed-source clawmetry-pro boundary depends on an unseen ENTITLEMENTS.md.
Failure messaging is strong: tests assert clean lock-conflict and decrypt errors, CI distinguishes setup-flakes from regressions, and the benchmark refuses to print unstable numbers. Deducted for self-consistency: the scope description says 26 runtimes while the README says 30 ('the other 27'), and package. version 0.0.1 does not reconcile with the PyPI release story. Dependency availability is asserted (minimal deps, certifi fallback) but not shown in code-level fallback paths.
Audience and scenarios are well split (workstation / shared host / self-hosted / air-gapped deployment guidance each given), capability boundaries are unusually honest (per-runtime visibility matrix, 'zero means blind', observation vs blocking), and environment fit covers three OSes, Docker, and Python 3.8+. Deducted for trigger precision: the built-in detectors ship enabled by default, but their rule logic is evidenced only by screenshots and lists, not source in the reviewed files.
Information architecture and install notes are thorough (i18n, docs table, pip/Docker/agent-skill installs), and the known-limitations section in SECURITY.md is high quality — loopback-no-credential trust, token-in-URL, missing SSO all stated plainly. MIT license is present and matches the copyright holder. Deducted for: no CHANGELOG in the reviewed files and versioning resting on a PyPI badge; naming instability (package. is 'openclaw-dashboard-repo'); maintenance responsibility has response targets that are explicitly non-contractual, and publisher Instalabs LLC is registry-unverified.
Marginal value is clear: zero-config reading of existing session files, per-provider window sizing, and the free/paid split are stated up front, with observation-vs-enforcement honestly distinguished. Cost-benefit is excellent: per-path overhead measured, the 12%-of-core figure over the project's own 5-10% budget published as a bug. Deducted for output usability: dashboard quality can only be inferred from screenshots and endpoint descriptions in a static review, and most of the 30-runtime coverage is carried by a closed-source companion, leaving only 3 runtimes fully verifiable in the open code.
Fact/inference separation is handled well: context-window provenance labels, a reproducible benchmark harness that lists what remains unmeasured, and the E2E page-integrity script that states its own limits ('protects you from a compromised CDN, not from the vendor'). Deducted for traceability: key claims (EGRESS inventory, compatibility matrix, entitlements, overhead data) point to files not included in this review, so only the references are confirmable; cross-source corroboration largely holds (README and SECURITY.md security models align; the test file backs encryption and no-plaintext-on-wire), but the 26/30 runtime discrepancy and an unexplained 'osai-verify' footer comment are minor unresolved points.
- An anonymous install ping, PyPI version check, and one public-IP lookup run by default; all claim to carry no session content and are opt-out, but read the egress inventory and disable as needed before installing.
- Loopback requests are trusted without a credential: on shared build hosts or multi-tenant environments any local process can read the dashboard API — run under a dedicated user or do not deploy there.
- 27 of 30 runtimes depend on the closed-source clawmetry-pro companion, whose behaviour cannot be statically verified from this repository; for high-sensitivity environments prefer the local free path or an air-gapped setup.
- No CHANGELOG is present and security fixes ship only as new releases with no LTS backport branch; teams pinning versions must track patches themselves.
- Approvals/gating and the enforcement proxy are off by default and paid-tier features; observation is not blocking — do not assume protection exists merely from installing.
- Publisher Instalabs LLC is registry-unverified; evaluate the trust boundary of the closed-source component and cloud sync against a self-hosted, air-gapped baseline.
What does this agent do, and when should you use it?
ClawMetry (vivekchand/clawmetry, MIT licensed) is a local-first observability and governance dashboard for AI agents. After `pip install clawmetry && clawmetry`, it auto-detects the agent runtimes already on your machine — 30 supported, including OpenClaw, NVIDIA NemoClaw, Claude Code, Cursor, OpenAI Codex, GitHub Copilot, Gemini CLI, Cline, OpenHands, and Aider — and tails their local session files read-only in a separate process, changing nothing about how they run. The dashboard at http://localhost:8900 shows sessions with turn-by-turn replay, per-runtime/model/session/day cost and token breakdowns, a live Flow message diagram, a Brain reasoning/tool-call event stream, per-provider context-window utilization, memory and skill loading, health and logs, and alerts routed to Slack, Discord, PagerDuty, Telegram, or Email. It can also pause risky tool calls before execution via an opt-in hook gate. Data stays on your machine by default; `clawmetry connect` uploads an end-to-end encrypted snapshot whose key never leaves your machine. The free tier covers only OpenClaw, NVIDIA NemoClaw, and Goose; the other runtimes, cloud sync, and governance features (approvals, policies, anomaly detection, OTel export) require paid plans ($9 or $19 per node per month).
Running clawmetry detects local agent runtimes and tails their session files and logs read-only from a separate process, rendering a dashboard on port 8900. Outputs include: turn-by-turn session transcripts with replay; cost and token data aggregated per runtime, model, session, and day with anomaly flags; a live Flow diagram; a Brain reasoning/tool-call stream; context-window utilization computed from a readable per-provider table in clawmetry/context_windows.py (Anthropic, OpenAI, Google, xAI, DeepSeek, Kimi, Qwen, Mistral, Llama, GLM), with GET /api/context-coverage reporting whether a zero means clean or blind per runtime; memory/skill load records; and disk/memory/error-rate health plus live logs. Optional paths: an HTTP interceptor (CLAWMETRY_INTERCEPT=1, +0.44 ms per LLM call) tracks agents built on SDKs; a pre-tool hook gate (~44 ms) and an enforcement proxy (+9.7 ms) enable pre-execution approvals and policies. Built-in detectors run from install: agent-offline, telemetry stopped, cost spike, token burst, error climb, error spike, budget threshold, threat signature, security findings — routed to Slack, Discord, PagerDuty, Telegram, or Email.
- A developer running several coding agents (Claude Code, Cursor, Codex) who wants one dashboard showing each agent's 24-hour and lifetime cost, last-seen time, and current work.
- A team lead who needs to review agent transcripts and per-tool timing without interrupting runs — for example triaging the 324 of 667 sessions the sample dashboard flagged with waste signals.
- A budget-constrained user who wants cost anomaly alerts (e.g., spend at 7x the daily average) routed to Slack or PagerDuty.
- A security-conscious team that wants risky tool calls (recursive deletes, force pushes, sudo, package installs) held for approval, including from a phone.
- A developer with a custom SDK-based agent who wants LLM call and cost attribution via the HTTP interceptor.
- A team needing audit compliance can enable Pro's tamper-evident audit log and OpenTelemetry export.
What are this agent's strengths and limitations?
- Genuinely zero-config: one command auto-detects and read-only tails every local agent runtime in a separate process, changing nothing about how agents run.
- Context-window sizing is per provider (Anthropic, OpenAI, Google, xAI, DeepSeek, Kimi, Qwen, Mistral, Llama, GLM) with provenance labels (model_table, explicit_marker, observed_floor, default) instead of one vendor's ruler for all runtimes.
- Transparent overhead: every instrumentation path has measured numbers (0 for session tailing, +0.44 ms/LLM call for the interceptor, +9.7 ms for the proxy), with an open benchmark harness run in CI across Linux, macOS, and Windows.
- Local-first privacy: no session data leaves the box by default; connect uploads are end-to-end encrypted, and the browser decryption code is byte-comparable to the published PyPI artifact.
- Most mainstream runtimes (Claude Code, Cursor, Codex, Copilot, etc.) are paid-tier only — the free app covers just OpenClaw, NemoClaw, and Goose; Starter is $9/node/month, Pro $19/node/month.
- Governance features (approvals, tool-risk policies, anomaly detection, OTel export, audit log) require the Pro plan, and risk-call gating ships fully off, so you must enable each rule.
- The docs admit the daemon sustains about 12% of one core, over the project's own stated 5-10% budget, and is published as a bug to chase; the enforcement proxy is ~7x slower on Windows than Linux.
- Compaction events are only visible on some runtimes, so context coverage can flag blindness but not fix it; SDK tracking of your own agents requires extra configuration (CLAWMETRY_INTERCEPT=1).
How do you install or deploy this agent?
Requires Python 3.8+ on macOS, Linux, or Windows, and at least one agent runtime on the same machine. Install and start: pip install clawmetry && clawmetry — the dashboard opens at http://localhost:8900. One-liner alternative: curl -sSL https://raw.githubusercontent.com/vivekchand/clawmetry/main/install.sh | bash. Docker instructions are in docs/DOCKER.md. Agents can set it up for you: npx skills add vivekchand/clawmetry --skill agent-kill-switch. Paid nodes activate a self-hosted license with clawmetry license.
How do you use this agent?
Run clawmetry after installing — no configuration needed. It auto-detects runtimes and opens http://localhost:8900; a header switcher re-scopes every tab to a selected runtime. To track your own agent, set CLAWMETRY_INTERCEPT=1 to enable the HTTP interceptor (docs/SDK_TRACKING.md). Approval and risk policies are off by default; enable each rule in the Approvals panel. Cloud sync requires clawmetry connect (end-to-end encrypted). Default behaviors — an anonymous install ping, a PyPI version check, and one public IP lookup — are opt-out and documented in docs/EGRESS.md. Reproduce overhead benchmarks yourself with python -m benchmarks.overhead.