agent-teams

MOAA: Mother of All Agents

Turn Claude Code into a team. MOAA gives you a Specialist Council of 60 expert agents, each with its own model and effort level, and an Agent Arena where agents compete and get judged. Agents that keep winning are forged into new permanent council members. For the biggest jobs, a War Room runs a whole team of parallel Claude Code sessions, with review gates and a live control app. Allies hands tasks to Codex, Gemini, Kimi or DeepSeek when you want another model family or need to save Claude usage, and Strata carries what matters from one session to the next. It installs with one command or loads as a plugin marketplace.

Created by Ahmad Hammoud. The Specialist Council and the Agent Arena are original work; everything else is credited in NOTICE and THIRD_PARTY.md.

Quick start

git clone https://github.com/ahmadhammoud56/MOAA.git && cd MOAA
./install.sh              # macOS / Linux / Git Bash  (Windows PowerShell: .\install.ps1)

Restart Claude Code, then try:

/specialist-council:specialist-council  Is 221 prime? Factor it.
/agent-arena-mode:judge --quick -- Write fib(n) with fib(1)=fib(2)=1 and give fib(20).

Prefer to pick plugins yourself? Run /plugin marketplace add ahmadhammoud56/MOAA inside Claude Code; see Use it as a plugin marketplace.

What a council run looks like

The animation replays a real run from its transcript, compressed from 44 s. Here is the coordinator's text output from an earlier run of the same task (lightly trimmed):

221 is not prime. 221 = 13 × 17.
Both workers agree. Check: 13 × 17 = 221, and 15² − 2² = 221. 13 and 17 are prime.

Dispatched (mode: ensemble, parent as judge):
- specialist-council:council-analyst                        claude-opus-5-5 / high    (Fermat factorization, gcd test)
- council-forged-independent-method-arithmetic-checker      claude-sonnet-5-5 / high  (trial division, difference of squares)
council.py resolve ran for both, with no permission prompt.

The forged checker in this run is an arena agent that was validated in 2 separate runs with scores of 97 and 95, then promoted into the council. That loop is what MOAA is built around:

 task ──► Specialist Council (60 roles, per-role model + effort) ──► verified answer
            ▲                                      │
            │ forge (≥2 validated runs, every      ▼
            │ score ≥ 85, spread ≤ 10)      Agent Arena (8 modes, judge + validator)
            └──────────── library of winning agents ◄─┘

Record your own

tools/council-replay turns any council session transcript into the animation above. It is optional, isn't installed, and needs only Python and Chrome, Edge or Chromium:

python tools/council-replay/record.py ~/.claude/projects/<project>/<session>.jsonl -o council-run.gif

record.py serves the page on 127.0.0.1, records it in a headless browser and writes the GIF. With no browser found it prints a URL instead; open it and click Export GIF. Run extract.py <session>.jsonl to see the replay data first. User paths, e-mail addresses and secret-looking strings are redacted, but check the GIF before you share it. To reproduce the demo, use tests/fixtures/council-run/session.jsonl.

The same task in an interactive session with Specialist Council 1.3.1: both workers were dispatched on the first try. The coordinator resolved only the forged member, so the analyst ran on its frontmatter defaults (tests/fixtures/council-run-interactive):

War Room

A team of real Claude Code sessions working one big job in parallel. Subagents share one context and one turn. A War Room is different: it starts several independent claude sessions, each with its own Specialist Council role, model, effort, plugins and permissions. They coordinate through a shared task board and message bus, review each other's work and merge it into one branch. A local app shows the whole team at work.

How a run works

1. Plan. Describe the job, either in the app (Home → New run) or to Claude in a chat (/war-room). The plan names the workers (for example a planner, two builders, a reviewer and an arbiter), the seed tasks and the budget. In the app, Draft with Claude writes the plan with one headless Sonnet call (about 5k tokens, no tools); a team template costs nothing. You can edit every worker and task before saving, and run any worker on an ally engine (Codex, Gemini, Kimi, DeepSeek or OmniRoute) instead of Claude.
2. Preview and launch. The preview lists every worker with its role, model, effort and plugins, plus a calibrated token estimate and the usage guard. Launching needs war mode armed, an acknowledgement and a confirm code tied to that exact plan.
3. Work. Workers claim tasks from the board, message each other and checkpoint as they go. Builders write in their own git worktree. Parked workers wait for new work without using tokens, and the conductor wakes idle ones when work arrives.
4. Review gates. Finished work counts only after a reviewer on a different model passes it (score 85 by default). Failed work goes back to its author. After repeated failures an arbiter decides, and after that you do.
5. Merge. Approved commits merge into a per-run branch, war-room/<run>. Merge conflicts go back to the author.
6. Finish. When everything is merged, the run stops its workers, measures the real token cost (this calibrates later estimates), removes its sessions and merged worktrees, and writes a report. Merging into your own branch is a separate Integrate step, and it never pushes.

The app

• Pages: Home (war mode switch, plan usage, quick-start templates, runs), New run, the run view, History and Settings, behind a slim sidebar. Ctrl+K opens a command palette everywhere.
• New run: a folder picker, a job prompt and team templates (feature team, bug hunt, review swarm, research fan-out, mixed allies, refactor) that fill the plan editor at no token cost. Each worker gets a role and an engine: a Claude tier, or an ally.
• Run view:
  • a needs-attention strip for human gates (approve or reject in place), failed reviews, spend near the cap and a silent conductor;
  • a KPI strip: tasks merged, the 5-hour window, spend, who is working (Claude versus allies) and gates waiting;
  • the team, coloured by model family (Claude, Codex, Gemini, Kimi, DeepSeek, OmniRoute);
  • Board (a kanban), Flow, Tree, Topology and Cost views, review gates, the merge queue, activity and a family-coloured timeline.
• Inspectors: click a task for its reviews, notes, commit and diff stat; click a worker for its tool calls, messages, checkpoints and usage charts, and a button to raise its console.
• Settings: team size, the usage guard, a spend cap, gates, presets, alerts, th