← Back to Discover

agentwatch

Local-only viewer that merges the logs of the AI coding agents on your machine into one live timeline, with per-turn cost, anomaly alerts, budget caps and an MCP server mode.

Observability & evalsCommand-line tools
Visit website ↗

What it does

agentwatch reads the session files that coding agents such as Claude Code, Codex, Gemini CLI, Hermes and OpenClaw write on your machine and merges them into one live timeline. Cursor is covered at the configuration level only. It runs locally with no cloud, no telemetry and no sign-in. You use a terminal view for live tailing and a browser interface on a local address for drilling into projects, sessions, token charts, call graphs and prompt replay. It works out per-turn token use and dollar cost with cache-aware rates, charts how full the context window gets and where compaction happened, and flags unusual cost, duration or stuck-loop patterns. You can set per-session and daily spend caps that raise desktop alerts, and define your own notification rules. It can also run as an MCP server so agents can query their own history, and export OpenTelemetry spans. It is a viewer rather than a background daemon, so it records events only while it is running, and it supports macOS and Linux only.

Capabilities

  • One live timeline across several coding agents
  • Per-turn token and cost accounting with cache-aware rates
  • Anomaly detection for cost spikes and stuck loops
  • Spend caps and custom notification rules
  • MCP server mode and OpenTelemetry export
  • Local browser interface with token charts, call graphs and replay

When to use it

  • You run more than one AI coding agent on the same machine and want a single view
  • You want an alert when an agent overspends or gets stuck in a loop
  • You need agent activity data that stays on your machine

Alternatives to explore

  • claude-devtools
  • Langfuse