Pick your team. Hand off the work.
Choose the coding agents brindle may run: Claude Code, Codex, Google Antigravity or open-weight models. Give the supervisor a goal and it splits the work into tasks, runs a worker for each on its own branch and git worktree, and merges a branch only after review and your checks pass.
A session
One tmux window: the brindle sidebar on the left, the supervisor's chat on the right. Simulated; drag to change how many workers run.
On Team, org policy can cap a member's parallel workers.
The supervisor splits it
Describe a goal. It becomes tasks, each sent to a model sized for it.
A reviewer pushes back
A reviewer checks every branch and sends weak work back to its worker.
Only passing work merges
With autopilot, a branch lands only once it's approved and your checks pass.
What else it does
A worktree per agent
Each agent gets its own branch, worktree and ports. Your setup script and .env run in each one. Merged worktrees and their branches are removed.
Milestone checks
Each milestone has a check command. Autopilot keeps going until they all pass or it needs you.
Conflict-aware merging
Least-overlapping branches merge first; conflicts go back to their worker.
Usage limits
Light tasks go to lighter models. Near your Claude limit, autopilot pauses Claude workers and restarts them when the window resets.
Undo a wrong turn
Rewind a worker to any turn and restart it with your note.
Pick up where you left off
Close the chat and the session pauses, with every branch and conversation kept. Start brindle and pick any open or paused session.
Sessions that share a repo
Run several sessions on one repo. Each sees the others' work; a task that clashes with another session's queues and starts from its merge.
Shared decisions
Record a decision once and agents in every session follow it.
A knowledge base of your docs
Agents search an index of your docs that stays on your machine. Secrets are redacted and injected instructions quarantined.
Other agent orchestrators vs. brindle
Other orchestrators give each agent a worktree, and agent CLIs' subagents can now split the work. brindle also runs it, gates it and works alongside them.
| Topic | Other orchestrators | brindle |
|---|---|---|
| A worktree per agent | ✓Each session gets its own branch and worktree. | ✓Each agent gets its own branch, worktree and ports. Your setup script and .env run in each one. |
| Agents | ✓Claude Code, Codex and others, depending on the tool. | ✓Claude Code, Codex, Google Antigravity and open-weight models, including local ones. |
| Pause and resume | ✓Sessions and workspaces stay until you remove them. | ✓Close the chat and the session pauses, with every branch and conversation kept. |
| Undo | ✓Some revert code and chat to an earlier turn; others rely on the agent CLI. | ✓Rewind a worker to any turn and restart it with your note. |
| Cost | ✓Some estimate cost per agent or show token totals. | ✓Tokens and cost per agent in the sidebar. Budgets per member on Team. |
| Splitting the work | ✓You split it across sessions, or an agent CLI's subagents split one task, nested a few layers deep. | ✓A supervisor splits your goal into milestones with check commands and starts a worker, on its own branch, for each task. |
| Picking models | −You pick the model for each session. | ✓Each task goes to a model sized for it, across providers. |
| Review | −You read the diff. | ✓A reviewer checks every branch, approves the exact commit, which can be another model, or sends weak work back to its worker. |
| Done means | −You decide. Some tools show status or CI. | ✓Each milestone's check command passes. Autopilot keeps going until they all pass or it needs you. |
| Merging | −You open the pull request, merge it and fix conflicts. | ✓With autopilot, a branch lands only after approval and passing checks. Least-overlapping branches merge first; conflicts go back to the worker. |
| Task order | −You start the sessions in the order you want. | ✓A task queues until the tasks it depends on have merged. |
| Usage limits | −You watch your limit and stop agents yourself. | ✓Autopilot pauses Claude workers near your limit and restarts them when it resets. |
Left to youDoes it
brindle enforces review and checks in its own client. To enforce them for every client, turn on branch protection in your Git host.
Get started
Step-by-step setup for each plan. Free runs everything on your machine and needs no account. Run brindle demo for the whole loop on a practice repo, or brindle init in your own.
brindle is free to use, on your own machine and at work: Brindle License 1.0 lets you install and run it for any purpose, but not modify it, redistribute it or use it to run a competing service.
Needs macOS or Linux, tmux, git and Claude Code. Codex or Antigravity can run the supervisor (--supervisor); a local model can run workers and reviewers.
For teams
Team and EnterpriseIdentity
SSO on Team- SSO through SAML or OIDC; can be required for members
- Custom role mapped to an Okta group (Enterprise)
- SCIM 2.0 provisioning and deactivation (Enterprise)
Cloud credentials
Team- Company routes on Amazon Bedrock, Google Vertex AI or Azure Foundry, federated through PawDelta
brindle loginissues credentials that expire within minutes; no long-lived cloud keys on laptops- Personal API keys can be disabled (Enterprise)
Budgets
Team- Per seat per month, per goal and per task; the most specific wins
- For everyone, a role or one member
- Throttle a member to cheaper models or fewer parallel workers
- Alerts at 80% and past 100% of the monthly budget