ClawMetry
ClawMetry is a real-time observability dashboard for AI agent runtimes, supporting 30 runtimes including OpenClaw, Claude Code, and OpenAI Codex. It provides session transcripts, cost tracking, flow diagrams, and alerts without modifying agent behavior. It is for developers and teams running AI agents who need visibility into performance, costs, and potential issues.
✨ Key features
- Sessions & transcripts with turn-by-turn replay
- Cost & token tracking per runtime, model, session, day
- Live flow diagram of messages through channels, models, tools
- Context blowout monitoring with per-provider window sizing
- Alerts routed to Slack, Discord, PagerDuty, Telegram, Email
- Approvals to pause risky tool calls before execution
🎯 Use cases
- Monitor multiple AI agents from a single dashboard
- Track and optimize token usage and costs across runtimes
- Debug agent behavior with session replays and flow diagrams
- Set up alerts for budget caps, error spikes, or agent downtime
- Implement approval workflows for risky tool calls
📦 Installation
🧰 Requirements: Python 3.8+ on macOS, Linux, or Windows, with at least one supported agent runtime installed locally.
pip install clawmetry # then: clawmetry
Or the one-liner: curl -sSL https://raw.githubusercontent.com/vivekchand/clawmetry/main/install.sh | bash
Needs Python 3.8+ on macOS, Linux or Windows, and at least one agent runtime on the same machine. Docker instructions: docs/DOCKER.md.
🚀 Usage
pip install clawmetry && clawmetry
Opens at http://localhost:8900. Zero config: it finds the agent runtimes you already have, reads them read-only, and changes nothing about how they run.
⚠️ Good to know
The daemon host cost is about 12% of one core, which is over the stated 5-10% budget and is published as a bug to chase. Also, context coverage may be blind on some runtimes, and the tool cannot see compaction events on all runtimes.
❓ FAQ
Which agent runtimes are supported?
ClawMetry supports 30 runtimes including OpenClaw, NVIDIA NemoClaw, Goose, Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Gemini CLI, and many more. Free runtimes include OpenClaw, NemoClaw, and Goose; others require a paid plan.
Does ClawMetry send my data to the cloud?
No. Session data stays on your machine unless you run clawmetry connect. When you connect, data is end-to-end encrypted with a key that never leaves your machine. Default installs send only anonymous pings and version checks, which are opt-out.
How much overhead does ClawMetry add to my agents?
Session-file tailing adds 0 overhead. The HTTP interceptor adds about 0.44 ms per LLM call, the pre-tool hook gate adds about 44 ms per gated call, and the enforcement proxy adds about 9.7 ms per LLM call. The daemon uses about 12% of one core.
How does ClawMetry handle context window differences between providers?
It uses a per-provider table of context window sizes, so utilization percentages are accurate. It also reports provenance for each window size and indicates when a zero means 'ran clean' or 'we're blind' via the context-coverage API.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.