For Claude Code and Codex

Arnheim

Arnheim showing Fernhollow, a made up sample company, as a building with project floors, daily token spending and budgets. Rooms glow warm where agents work and amber where something needs you. A panel lists seven things needing you, including command permission with Allow and Deny.

AI coding agent dashboard: Arnheim shows which Claude Code or Codex agent is waiting on you and whether work called done passes checks.

An open source tool from Allwhen Labs.

Run Arnheim

Run this command in your terminal.

git clone https://github.com/allwhenlabs/arnheim && cd arnheim && python3 serve.py

Go a little deeper.

The workings, the measurements and the things that needed fixing.

Start with how a tool works. Keep reading to see what supports it, what broke and why we built it.

ProvenWe measured it or a test shows it.

PlausibleStated, not yet measured by us.

A guessNot tried yet.

Inside Arnheim

The agent dashboard

How it works

Plausible

Arnheim shows your AI coding agents as a building you can walk through. What needs you is listed, most urgent first.

Work an agent calls done is marked claimed until its checks pass.

Each floor has a lead who knows that project, and you can talk to them by typing or voice. Pass on an instruction and they tell you next time how it went, without being asked.

What we measured

  • Proven

    312 tests, all passing when run on 17 Nov 2026.

  • Proven

    A restore point gets at most 20 seconds a step and 40 seconds in all. A folder that takes longer gets no restore points for six hours.

  • Plausible

    A walk through the building loads about 6 MB, once, when someone presses Walk.

  • Plausible

    Voice needs about 1.2 GB of disk. Its models download once, about 470 MB.

  • Proven

    93 changes between 8 Nov and 17 Nov 2026.

What broke

  • Proven

    Pressing two keys quickly sent two answers to one permission request, and the screen showed a confusing note. Fixed 15 Nov: each request is answered once.

  • Proven

    A time limit could cut in between saving a restore point and keeping it, and code files named like api_key_utils.py were wrongly treated as secrets. Fixed 12 Nov.

  • A guess

    Linux and Windows should work but haven't been tried.

Why we built it

Plausible

To see at a glance which of your AI coding agents is waiting on you, and whether the work each one calls done really passes its checks.

Try Arnheim.

Tests and builds need somewhere to run? Meet Orchard.