Albert

Albert

Give Claude Code a goal and walk away. A roster of specialized agents plans it, splits
the work, builds it in parallel, reviews and QAs its own output, and merges when the
critics sign off, while a live high tech HUD shows you every agent as it works.

This repo is three cooperating pieces:

- Albert is a long-running autonomous harness for Claude Code.
  You give it a goal in plain English. An orchestrator (A.L.B.E.R.T., the Agentic Loop
  Broker for Execution, Reasoning & Tasking) decomposes it into tasks, assigns each to the
  cheapest model that can do it, fans them out concurrently in isolated git worktrees, runs
  independent verify / code-review / QA critics on every result, and merges the work when
  and only when those critics pass. It keeps going, iteration after iteration, until the
  goal is met or it hits a checkpoint you asked for.
- The Albert Console is a zero-dependency local dashboard that watches all of the
  above in real time, across every project on your machine, in a cyan-and-gold "agent OS"
  HUD.
- Albert Chat (optional) is a direct line to the orchestrator, right inside the console:
  a fly-up chat dock where a read-only concierge answers questions about any run in seconds,
  steering messages queue into the running orchestrator's inbox, and A.L.B.E.R.T.'s replies
  come back into the conversation. It can also launch new runs for you.

Everything runs locally. The console is read-only and binds to 127.0.0.1 only.

See it

Fleet — every agent as a card, grouped by role, with its model tier, live status, and
heartbeat.

Comms — a live feed of who dispatched what to whom (A.L.B.E.R.T. -> loop-worker,
loop-verifier-dev -> A.L.B.E.R.T.), streamed over Server-Sent Events.

Sessions — the full history of every run across every project: goal,…