# brat

> brat is a multi-agent harness for AI coding tools. It coordinates Claude Code, Aider, OpenCode, Codex, Continue, Gemini, and GitHub Copilot working in parallel on the same codebase, with crash-safe state built on the Gritee append-only event log. Distributed as a single Rust CLI under the MIT license by neul-labs.

brat is an open-source project. There is no SaaS, no account, no telemetry. The CLI is the unit of correctness; an optional HTTP daemon (`bratd`) auto-starts when the web dashboard or multi-session coordination is needed, and idle-shuts-down after fifteen minutes by default.

## Docs

- [Home](https://brat.neullabs.com/): Overview, roles, supported engines, architecture
- [Features](https://brat.neullabs.com/features/): Crash-safe WAL, seven engine adapters, actor isolation, TTL leases, the Refinery merge queue
- [How it works](https://brat.neullabs.com/how-it-works/): The four-layer architecture and how a plan becomes a merged branch
- [Quickstart](https://brat.neullabs.com/quickstart/): Install, init, and run your first convoy of agents in five commands
- [Glossary](https://brat.neullabs.com/glossary/): Mayor, Convoy, Task, Witness, Refinery, Deacon, Gritee, WAL, and more
- [About](https://brat.neullabs.com/about/): Why brat is a CLI, not a SaaS; the design thesis
- [FAQ](https://brat.neullabs.com/faq/): Crash-safe parallel agents, resuming after a crash, orchestrating Claude Code + Aider + Codex
- [Docs subdomain](https://docs.neullabs.com/brat/): Full technical documentation
- [Source](https://github.com/neul-labs/brat): MIT-licensed Rust workspace

## Use cases

- [Use cases index](https://brat.neullabs.com/use-cases/): Bug-triage fleets, engine bake-offs, overnight refactors, CI-gated merge queues
- [Parallel bug triage across a codebase](https://brat.neullabs.com/use-cases/parallel-bug-triage/): Point the Mayor at src/, let it file one task per bug, and run a fleet of agents to fix them at once
- [Running a multi-engine bake-off](https://brat.neullabs.com/use-cases/multi-engine-bake-off/): Give the same task to Claude Code, Aider, and Codex at once and compare the branches side by side
- [Overnight refactor fleet](https://brat.neullabs.com/use-cases/overnight-refactor-fleet/): Queue a large mechanical refactor as dozens of tasks and let the fleet grind through it while you sleep
- [A CI-gated agent merge queue](https://brat.neullabs.com/use-cases/ci-driven-agent-merge-queue/): Let agents propose branches all day, and land only the ones that pass your existing CI, in order
- [Graduating from tmux panes](https://brat.neullabs.com/use-cases/graduating-from-tmux/): Move from counting tmux panes to querying an event log once the fleet outgrows your screen
- [Auditable, reproducible agent runs](https://brat.neullabs.com/use-cases/auditable-agent-runs/): Keep an immutable record of every task, session, lock, and merge for review and reproducibility

## Blog

- [Blog index](https://brat.neullabs.com/blog/): Field reports on running agent fleets, architecture deep-dives
- [Running 50 Claude Code agents in parallel: what breaks first](https://brat.neullabs.com/blog/running-50-claude-code-agents-in-parallel/): When you scale from one coding agent to fifty, the bottleneck is the orchestration layer — what failed in our experiments and how brat responds
- [Designing crash-safe state for ephemeral agent processes](https://brat.neullabs.com/blog/crash-safe-state-for-ephemeral-agent-processes/): AI coding agents crash a lot; treat their state as durable and you lose it. How brat's append-only WAL design handles it
- [Why a thin CLI beats SaaS for agent orchestration](https://brat.neullabs.com/blog/why-a-thin-cli-beats-saas-for-agent-orchestration/): The design reasoning for a Rust CLI you install and forget instead of a hosted control plane, including the uncomfortable parts
- [Mayor, Witness, Refinery, Deacon — the six-role vocabulary](https://brat.neullabs.com/blog/the-six-role-vocabulary-mayor-witness-refinery-deacon/): What each orchestration role does, why it exists, and what it explicitly does not do

## Compare

- [vs Claude Squad](https://brat.neullabs.com/compare/claude-squad/): Claude Squad supervises a handful of agents in worktrees + tmux; brat is the durable layer below for driving larger fleets crash-safely
- [vs AutoGen](https://brat.neullabs.com/compare/autogen/): AutoGen is a Python library for in-process conversational agent graphs; brat is a Rust CLI running external coding agents on a real repo
- [vs GNU parallel](https://brat.neullabs.com/compare/gnu-parallel/): GNU parallel fans out invocations but loses state, audit trail, lock discipline, and a merge queue
- [vs raw tmux + scripts](https://brat.neullabs.com/compare/raw-tmux/): The honest baseline — brat replaces the bash with an event log once you outgrow counting tmux panes

## Optional

- [RSS](https://brat.neullabs.com/rss.xml): Blog feed
- [llms-full.txt](https://brat.neullabs.com/llms-full.txt): Long-form summary with verified facts