Multi-Agent Workbench
Run Claude Code, Codex and Antigravity in parallel. Each agent gets its own git worktree and task queue; you get real commits, check results and a live dashboard.
The live dashboard that opens when you run maw.
Install
npm install -g multi-agent-workbench cd your-project maw plan PLAN.md # split your plan into micro-tasks and assign agents maw doctor # agents installed and logged in? maw # run; live dashboard opens in your browser maw review # what each agent did maw merge --check "npm test" # merge finished work into maw/integration
Warning: agents run with permission prompts disabled inside their worktrees. Use it only where that is acceptable, and review every commit before merging.
What it does
maw planturns your plan into small tasks, each with an owner, a prompt, a check and dependencies- One queue and worktree per agent; queues run in parallel, isolated tasks alongside
- Failed checks go back to the agent, which continues its session and tries again
- Live dashboard: each agent's output, current task and what's next, every commit with its diff, changed files and cost
maw review,maw mergeandmaw cleanto integrate and tidy up- Claude Code, Codex, Antigravity, or any CLI you define
- Budgets, environment allowlists, sandbox wrappers and notifications
tasks.json
{
"plan": { "title": "Orders feature" },
"tasks": [
{ "id": "L1", "owner": "claude", "task": "Orders API", "retries": 1,
"prompt": "Add GET /api/orders. Run tests and commit.", "check": "npm test -- orders" },
{ "id": "C1", "owner": "codex", "task": "Orders page",
"prompt": "Build the orders page. Run lint and commit." },
{ "id": "A1", "owner": "antigravity", "command": "agy", "depends_on": ["L1", "C1"],
"prompt": "Merge maw/claude and maw/codex, run all tests, fix clear bugs, commit." }
]
}
Commands
maw plan [PLAN.md] | Split a plan into micro-tasks and write tasks.json |
maw doctor | Check agents are installed and logged in (exit 1 if not) |
maw run | Run and open the dashboard (--dry-run, --resume, --max-parallel, --notify) |
maw status / maw review | Summarize the last run / each task's commits and changes |
maw merge | Merge finished work into an integration worktree and test it |
maw clean | Remove maw's worktrees and branches |
maw dashboard | Reopen the dashboard for a finished run |
maw install-skill | Install the coordinator skill for Codex and Claude Code |
Use as a skill or plugin
/plugin marketplace add ruhul0/multi-agent-workbench /plugin install multi-agent-workbench@multi-agent-workbench
Or maw install-skill, then in Codex: Use $multi-agent-workbench to run PLAN.md across Codex, Claude and Antigravity. In Claude Code: /multi-agent-workbench.