Run a team of coding agents on your own machine and stay in charge of it. You talk to one director; it splits the job across sub-agents in their own git worktrees, and nothing counts as done until an independent verifier has seen it work.
It drives the Claude Code and Antigravity (Gemini) CLIs you already have, so each provider keeps its own login and subscription.
Download · Report a bug · Documentation
- Start a director for a repo. It is the one agent you talk to.
- It splits the work and hands each unit to a sub-agent with
spawn_subagent(playbook, brief, done_when). Code is written in a worktree on branchorch/<id>; your checkout is never touched. - A fresh verifier checks
done_when: a command and its output, a flow in the running app. FAIL goes back to the builder with findings; PASS tells the director which branch to merge. - You step in only when needed. Approvals, questions and failed checks land in one inbox.
Agents
- Live terminal cards for Claude Code and Antigravity on one canvas, with model and reasoning
effort per agent. Gemini models are read from
agy models, so new ones show up on their own. - Agents hand work to each other with
send_signal; every run is queued, capped and logged. - Role playbooks that fill themselves in from the project.
- A VS Code style editor with Vim modes on every agent's project, with Source Control across all its git repos. Nothing to install.
- Workspaces (isolated folders and agents) in groups, two side by side.
- Guard rails: approval gate, ping-pong and daily caps, kill switch, one prompt at a time per agent, reachable from your own machine only by default.
Director and sub-agents
- Playbooks:
feature,bugfix(reproduce first),refactor(record behaviour first),investigate(read-only). - A verdict is tied to the commit that was checked; a verifier that edits files has its verdict thrown away. Repeated FAILs stop at a cap and tell the director.
- Team tab: a pipeline board (Building → Verifying → Passed / Failed) with each sub-agent's step, verdict and diff; the director's decision log; who may message whom; the effort each task ran at; Reflect, which turns failed runs into a playbook change you approve.
- Sub-agents and verifiers always run Claude Opus 5.5. A FAIL steps the builder's effort up one level (to xhigh at most).
Staying in control
- Home inbox: every message waiting for approval, every question and every failed check across workspaces, with Approve / Deny in place.
- Activity: one timeline of messages and runs, with live transcripts.
- Ctrl/⌘K: jump to any agent or workspace, or run any action.
At least one, installed and logged in:
| CLI | Install |
|---|---|
| Claude Code | code.claude.com/docs |
| Antigravity CLI | antigravity.google/docs/cli — the agy command |
Windows (prebuilt). Download agent-orch-windows-x64.zip from Releases.
Unblock it before unzipping, or SmartScreen flags every file inside:
Unblock-File agent-orch-windows-x64.zipUnzip it and run agent-orch.exe inside the folder. Keep the folder together. The console window
that opens is the server: close it to stop.
Windows (from source). PowerShell, with Python 3.12 (winget install Python.Python.3.12) and
Git (winget install Git.Git):
git clone https://github.com/getAINow/donezo.git
cd donezo
py -3.12 -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python session_orchestrator.py servemacOS. The python3 built into macOS is 3.9, which is too old. Use Homebrew's:
brew install python@3.12 git
git clone https://github.com/getAINow/donezo.git
cd donezo
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python session_orchestrator.py serveLinux (Debian, Ubuntu; Python 3.10 or newer):
sudo apt install python3 python3-venv git
git clone https://github.com/getAINow/donezo.git
cd donezo
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python session_orchestrator.py serveThen open http://localhost:8992.
In the dashboard, open Settings › Connections and press Connect next to signal. It
carries messaging between agents and the director's team tools (spawn_subagent,
submit_verdict, record_decision…). Agents already running pick it up after a restart.
Or by hand, once per CLI:
claude mcp add --transport http --scope user signal http://127.0.0.1:8992/signal/mcp
agy mcp add signal http://127.0.0.1:8992/signal/mcpPut these in a .env file next to session_orchestrator.py (or next to agent-orch.exe):
| Variable | Default | When to set it |
|---|---|---|
ORCH_PORT |
8992 |
Port 8992 is taken |
CLAUDE_BIN / ORCH_AGY_BIN |
claude / agy |
The dashboard cannot find an installed CLI |
ORCH_HOST |
127.0.0.1 |
Only to reach it from another machine |
ORCH_NVIM_BIN |
empty (download) | The editor's Neovim download fails, or you want your own nvim 0.12+ |
ORCH_SUBAGENTS_MAX |
4 |
Active sub-agents per director |
ORCH_SUBTASK_MAX_ATTEMPTS |
3 |
Failed verifications before a sub-agent's work is marked failed |
Warning
Agents run shell commands with permissions bypassed, and there is no API key.
ORCH_HOST=0.0.0.0 hands that to everyone on your network.
Run it as your own user, never root. Replace the paths with where you unzipped or cloned it.
Windows — Task Scheduler
$exe = "$HOME\agent-orch-windows-x64\agent-orch.exe"
Register-ScheduledTask -TaskName "Donezo" `
-Action (New-ScheduledTaskAction -Execute $exe -WorkingDirectory (Split-Path $exe)) `
-Trigger (New-ScheduledTaskTrigger -AtLogOn) `
-Settings (New-ScheduledTaskSettingsSet -StartWhenAvailable `
-RestartCount 3 -RestartInterval (New-TimeSpan -Minutes 1))Remove it with Unregister-ScheduledTask -TaskName "Donezo".
macOS — launchd
cat > ~/Library/LaunchAgents/com.donezo.plist <<PLIST
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0"><dict>
<key>Label</key><string>com.donezo</string>
<key>ProgramArguments</key><array>
<string>$HOME/donezo/.venv/bin/python</string>
<string>$HOME/donezo/session_orchestrator.py</string>
<string>serve</string>
</array>
<key>WorkingDirectory</key><string>$HOME/donezo</string>
<key>RunAtLoad</key><true/>
<key>KeepAlive</key><true/>
<key>StandardOutPath</key><string>$HOME/Library/Logs/donezo.log</string>
<key>StandardErrorPath</key><string>$HOME/Library/Logs/donezo.err</string>
</dict></plist>
PLIST
launchctl load -w ~/Library/LaunchAgents/com.donezo.plistStop it with launchctl unload -w ~/Library/LaunchAgents/com.donezo.plist.
Linux — systemd user service
mkdir -p ~/.config/systemd/user
cat > ~/.config/systemd/user/donezo.service <<'UNIT'
[Unit]
Description=Donezo
After=network-online.target
Wants=network-online.target
[Service]
Type=simple
WorkingDirectory=%h/donezo
ExecStart=%h/donezo/.venv/bin/python %h/donezo/session_orchestrator.py serve
Environment=PYTHONUNBUFFERED=1
Restart=on-failure
RestartSec=3
TimeoutStopSec=15
[Install]
WantedBy=default.target
UNIT
systemctl --user daemon-reload
systemctl --user enable --now donezo
loginctl enable-linger "$USER" # start at boot, not only when you log inLogs: journalctl --user -u donezo -f.
From source: git pull, then pip install -r requirements.txt, then restart. Prebuilt: download
the new zip and replace the folder, keeping your .env.
python verify.py runs every check_*.py (CI runs the same command). Each check guards a bug
that would otherwise fail silently — add one with any change that could.
MIT — see LICENSE.



