Agent Console shows every Claude Code and Codex session's tokens and cost, on this machine and any you connect

A week ago you could not tell which teammate was burning through the Claude Code budget. LockedIn Labs published Agent Console on 20 September 2026 and it turns that question into a screen. The repository has picked up 413 stars and 127 forks in seven days.

Agent Console is a local program that reads the transcripts Claude Code and Codex already write on your computer, then groups every session's tokens and estimated cost by model, by machine and by person. It runs on macOS, Linux and Windows, needs no account, and opens as a page in your browser signed in. The current release is v0.4.1.

What it prints

Under Tokens, last 24 hours, the console shows a total across every machine you have joined, the list-price estimate, the number of API responses (not transcript lines), and the split into cache read, cache write, output and uncached input, each with its share of the total. You can flip to the last hour, seven days or 30 days. A machine that has stopped reporting is called out on the chart, from when it went silent.

Under By model and By machine, each model's share and each machine's share appear as a row with tokens and dollar estimate. The list uses real Anthropic and OpenAI marks. Under Attention, the console names the one reading that needs it: a burn spike, spending without a session, a cache-write ratio that is off.

What travels off a machine

When you join a laptop or a build server to a shared team hub, LockedIn Labs says the wire carries only model ids, minute timestamps, token counts and salted hashes. Never a prompt, a reply, a file path or a file's contents. A test in the project pushes transcripts full of planted canary strings through the real reporter and hub and checks every byte that crosses the wire. Three opt-in switches send more if you turn them on: the first sends each project folder's name (never its path), and two others send alerts and tool activity as kinds and counts.

The console itself sends nothing anywhere. No telemetry, no update check, no account. That is the project's own statement.

Presenting a shared screen

Press P and every project, branch, machine and person becomes a stand-in name, estimates step back, and the strip reads PRESENTING. Press P again to stop. LockedIn Labs positions it for screen shares with people who should not see the client names.

Teams that already collect their own usage numbers through a gateway or metrics pipeline can turn on a matching input, so those readings show in a separate panel and the same request is not counted twice. LockedIn Labs names Claude Code's own telemetry, Kong and LiteLLM as the inputs it accepts today.

Two limits to know before you install. There is no public npm package yet, so today you use a source checkout, a GitHub release archive, or one of the release's standalone downloads for macOS, Linux and Windows. And the console reads Claude Code and Codex today; other coding agents are not read.

That still puts a shared, itemised, per-person AI spend number on one screen without shipping any transcript off your machines. Coding-agent spend has been the least readable line item in most engineering budgets this year, and asking staff to volunteer estimates was never going to give the finance team a real answer.

Source

• Repository: LockedinLabs-AI/agent-console on GitHub
• Release notes: v0.4.1