Once one coding agent is useful, the next step is obvious: run three. One refactors, one writes tests, one chases a bug. Then reality hits. They edit the same files, you lose track of which terminal is waiting for an answer, an SSH drop kills an hour of work, and nobody reviews what the agents wrote.
A new group of open-source tools fixes exactly this. Three of them cover most needs:
- Orca: a desktop app that runs several CLI agents side by side, each in its own git worktree.
- herdr: a terminal runtime that keeps agent sessions alive in the background and tells you which one is blocked.
- Squadra: a coordination toolkit where Claude Code, Codex and Cursor windows share tasks, and the window that wrote the code never reviews it.
Setup commands below come from each README. For the full comparison, see our best AI repos for running several coding agents in parallel page.
Quick answer
| Your problem | Use |
|---|---|
| Agents step on each other's files; you want to compare results | Orca |
| Agents run on a remote box over SSH and die when you disconnect | herdr |
| You need agents to hand off work, review each other and prove each step | Squadra |
1. Orca: parallel worktrees in one window
Orca's README describes it as a way to "run Codex, Claude Code, OpenCode or Pi side by side, each in its own worktree, tracked in one place." It works with any CLI agent: if it runs in a terminal, it runs in Orca.
The core idea is git worktrees. Each agent gets its own checkout of the repo on its own branch, so two agents can never overwrite each other's edits. The README's headline workflow: fan one prompt across five agents, compare the results, and merge the winner.
Install on macOS with Homebrew. Note the full cask name; the README warns that bare orca is an unrelated cask:
brew install --cask stablyai/orca/orca
Upgrade later with:
brew upgrade --cask stablyai/orca/orca
On Arch Linux:
yay -S stably-orca-bin
Windows and other platforms download a build from the releases page or onorca.dev.
Features that matter once you run several agents:
- Notifications and unread state, so you know which agent finished or needs attention.
- Usage tracking and an account switcher that shows Claude and Codex usage and rate-limit resets.
- SSH worktrees to run agents on a stronger remote machine.
- Annotate AI diffs: comment on any diff line and send it back to the agent.
-
The Orca CLI, so agents can script Orca itself with commands like
orca worktree create.
We break down that last one in our script agent workflows with the Orca CLI card, and the rest on the main Orca use cases page.
Orca is MIT licensed.
2. herdr: agents that survive your terminal closing
herdr calls itself "the runtime your coding agents live on." It is a single Rust binary, no Electron, that runs inside whatever terminal you already use. It does not wrap or replace Claude Code, Codex, Cursor or OpenCode; it owns their terminals.
Install it:
curl -fsSL https://herdr.dev/install.sh | sh
Or use brew install herdr or mise use -g herdr. Windows has a PowerShell installer in the README.
Then start it in your project folder:
herdr
Run your agents in panes, split as needed, and walk away. ctrl+b q detaches; running herdr again reattaches. If you know tmux, the prefix keys will feel familiar, and mouse click, drag and split work too.
What sets it apart:
- Detach without stopping work. Terminals keep running in a background server when you close the client or lose SSH. After a reboot, herdr restores the saved layout and can resume supported agent sessions (the README is clear that the original processes do not survive a restart).
- Every pane is marked working, blocked or idle. When an agent stops and needs an answer, herdr tells you, so you are not scanning ten panes for the stuck one.
- Several machines in one window, with a combined agent list across local and saved SSH machines.
- Agents can drive herdr through its CLI and socket API: spawn panes, prompt each other and wait until another agent is actually blocked.
If your main pain is losing work on remote servers, start with our keep agents running after SSH drops card. herdr is Apache 2.0 licensed.
3. Squadra: a team with rules, not just more windows
Orca and herdr manage where agents run. Squadra manages how they work together. It gives several Claude Code, Codex and Cursor windows one shared registry of tasks, messages, checks and decisions, all stored locally. Its rules are strict on purpose:
- Whoever writes a change does not review it.
- A task starts only after the recipient confirms it.
- Proof is bound to the exact code; if the code changes after the check, the proof no longer counts.
It needs Python 3.12+ and Git, and is tested on macOS and POSIX systems. The quickest install is inside Claude Code:
/plugin marketplace add dagoSte/squadra
/plugin install squadra@squadra
For Codex, in a terminal:
codex plugin marketplace add dagoSte/squadra
codex plugin add squadra@squadra
If you want the squadra command on your PATH, use the standalone installer, which keeps a backup so you can roll back:
git clone https://github.com/dagoSte/squadra.git
cd squadra
python3 -B scripts/install.py \
--backup "$HOME/.local/state/squadra/install-backup-001"
export PATH="$HOME/.local/bin:$PATH"
squadra --version
You mostly talk to it in plain words. Start a Claude window as coordinator with a prompt like the README's: "Use squadra for this project. Use the registry at /path/to/registry. Find the other windows and introduce yourself." Open the other windows on the same registry, give the coordinator a first task with a way to check it, and watch progress in the local dashboard:
squadra dash --root /path/to/registry --open
Each task moves through confirmation, work, an executed check, a review by another window and closure by the coordinator. It also tracks tokens per window and model across Claude Code and Codex. Squadra is MIT licensed with no account needed.
How to choose
- Solo developer, one laptop, wants to try several approaches at once: Orca. Worktrees solve the collision problem and the compare-and-merge flow is the fastest way to get value.
- Work happens on a remote server or you keep many long sessions open: herdr. It is the lightest option and it fixes the "which one is stuck" problem.
- Changes that need review and an audit trail, such as a team codebase: Squadra, possibly on top of either of the above.
They are not mutually exclusive. herdr inside an SSH session plus Squadra for the review rules is a reasonable stack.
Before you scale up
- Give every agent its own branch or worktree. Shared working directories are the number one cause of lost work.
- Watch usage. Three agents burn three times the quota, and Orca and Squadra both show it.
- Keep a human merge step before anything ships.
Get one email every Monday with the week's hottest AI repos and exactly how to use them: subscribe free at dedupedai.com
Top comments (0)