Monitoring

What your AI work costs, which AI processes are running, and what your tabs talk to.

⌥⌘U opens Monitoring. Everything here is computed on your machine; nothing is uploaded.

AI usage#

Cost, calls, cache hit rate and tokens, broken down by model, project and tool. The numbers come from the transcript files the agent CLIs already wrote on your disk.

Counting is not as simple as summing lines. Claude Code writes one JSONL line per content block, so a single API response appears several times with the same message id; those are de-duplicated by that id. Codex writes one usage record per response instead, keyed by response id. Agentty follows each format so a response is counted once.

Claude model prices are built in. For other models — the ones Codex uses, for instance — add prices in ~/.agentty/pricing.json and they are included:

json
{
  "some-model": { "input": 1.25, "output": 10.0, "cacheRead": 0.125 }
}

Prices are USD per million tokens. cacheRead, cacheWrite and cacheWrite1h are optional and default to 10%, 125% and 200% of the input rate.

Note

These are estimates for your own orientation, computed from local files. Your provider's invoice is the real number.

Processes#

The AI processes running on your machine right now, with what started them. Useful when something is still burning tokens after you thought it stopped.

Proxy#

A capture proxy that lists what your tabs talk to: start capture, open a tab, filter by endpoint.

  • Tabs opened while capture is on are routed through a proxy running on your own machine.
  • HTTPS stays encrypted. Only the host, port, byte counts and timing are recorded — no certificate is installed and nothing is decrypted.
  • For plain HTTP, the method, path and status are recorded, never headers or bodies.
  • Records live in memory, are limited in number, are never written to disk, and disappear when Agentty quits.

Stopping capture stops the recording. Tabs that were given the proxy keep using it so they don't lose their network.