What your AI coding agents actually cost, counted off your own disk
Agent Console turns the transcripts Claude Code and Codex already write to disk into per-session token, cache and cost dashboards — without a byte of prompt text leaving the machine.

What it does Agent Console reads the session transcripts Claude Code and Codex leave on your machine and renders them as a live dashboard: tokens split into cache read, cache write, output and uncached input; one lane per session with its model and subagents; burn rate in tokens per minute and dollars per hour; and each machine and person’s share of the total. Machines can join a self-hosted team hub, so laptops, build boxes and servers add up in one view — each keeps its own lanes and nothing is counted twice. It runs as a local Node process serving the console in your browser: no account, no build step, no dependencies.
The interesting bit The privacy claim is tested, not just asserted — a canary test pushes transcripts seeded with planted secrets through the real reporter and the real hub, then checks every byte that crosses the wire; only model ids, minute timestamps, token counts and salted hashes should survive. The numbers also have a conscience: estimates are labeled as estimates, and unknown readings display as unknown instead of quietly becoming zero, which already puts it ahead of most FinOps dashboards.
Key highlights
- Local-first: reads the agents’ own transcripts; no telemetry, no update check, no crash reporting
- Self-hosted team hub: every machine joins, keeps its own lanes, nothing double-counted
- Optional ingest of Claude Code OpenTelemetry and Kong or LiteLLM token metrics, plus a Prometheus
/metricsendpoint and a Grafana dashboard - Presenting mode (P) swaps project, machine and person names for stand-ins — screenshots safe by default
- Verifiable supply chain: Apple-signed and notarized macOS builds,
SHA256SUMS, signed attestations, and a CycloneDX SBOM from 0.4.1
Caveats
- Costs are list-price estimates, not reconciled against your actual bill
- Coverage is Claude Code and Codex only; other agents aren’t read
- Distribution is uneven: the npm registry package is unpublished, the Homebrew tap and container image are unverified, and the Windows executable isn’t code-signed
Verdict If your team burns tokens through Claude Code or Codex and wants cost and usage visibility without shipping prompt data to a vendor, this is a rare combination of useful and trustworthy. If you need other agents covered or real billing numbers, it isn’t there yet.
Frequently asked
- What is LockedinLabs-AI/agent-console?
- Agent Console turns the transcripts Claude Code and Codex already write to disk into per-session token, cache and cost dashboards — without a byte of prompt text leaving the machine.
- Is agent-console open source?
- Yes — LockedinLabs-AI/agent-console is open source, released under the MIT license.
- What language is agent-console written in?
- LockedinLabs-AI/agent-console is primarily written in JavaScript.
- How popular is agent-console?
- LockedinLabs-AI/agent-console has 801 stars on GitHub.
- Where can I find agent-console?
- LockedinLabs-AI/agent-console is on GitHub at https://github.com/LockedinLabs-AI/agent-console.