Brigade Stack is an open-source coordination layer for AI coding agents. It sits beside Codex, Claude Code, Cursor, OpenCode, T3 Code, and the rest, and gives them shared tasks, shared memory, and verification records any agent can check.
Stable: v0.30.0 on PyPI and npm. Beta: main is the 0.31 preview line. Channels
Why Brigade Stack
You run several agents at once. Sol in Codex takes one task, Opus manages another, and Argon takes a third here and there, whether they share one window in a multi-harness tool like T3 Code or each run in its own app. Your Grok Bot runs the rest of your life, and your Dots spin up work of their own. Every agent starts from its own conversation. None of them knows who owns which task, what the others already tried, or whether a fix was actually checked.
Brigade Stack holds that state outside any one tool. Agents read the same brief, claim the same tasks, and leave handoffs and evidence the next agent can inspect.
- Keep your tools. Works alongside the agents you already use. Start with one tool in one repository and add more when you need them.
- Work agents can pick up. Tasks, dependencies, claims, and handoffs make the next action discoverable without replaying a conversation.
- Checkable results. Verification records keep the command, exit code, logs, and Git state, so another agent can confirm "tests pass" for itself.
- You set the limits. Model access, scopes, budgets, and approval gates apply to Brigade-managed work, and you can inspect or stop it.
How it works
- Start with context. Agents begin from a work brief: relevant memory, open tasks, and existing evidence.
- Claim the work. Ownership is explicit, and conflicting or dependent work is visible.
- Run within your rules. Managed execution uses your configured models, scopes, budgets, and approvals.
- Verify the result. Checks run through Brigade and leave a record another agent can inspect.
- Hand off. The result, remaining blockers, and evidence stay where the next agent will find them.
A local setup needs no server. Fleet Hub adds shared ownership, presence, messaging, and model policy across machines when you want it. See the execution model.
Quick start
Install the stable bridge with npm (Node.js ^22.15.0 || >=23.11.0):
npm install --global [email protected]
brigadestack --version
In a Git repository you control, preview the setup, apply it, and read the brief:
brigadestack operator quickstart --target . --harnesses codex --dry-run
brigadestack operator quickstart --target . --harnesses codex
brigadestack work brief --target .
Replace codex with the harnesses you use, such as codex,claude. Record a first check:
brigadestack work verify run --target . --command "git diff --check" --capture brigade-work
This checks whitespace in your unstaged diff and records the command, exit code, and logs. Use your project's test command for subsequent checks. The first 10 minutes walks through health checks and receipt inspection.
The npm bridge exposes brigadestack and downloads a managed Python on first launch. Git is required for this walkthrough. For Python installs, platform limits, optional engine setup, and existing brigade installs, see Quickstart. Existing commands, Python imports, and .brigade/ data remain compatible.
Generated agent instructions and CLI suggestions still use brigade. On a bridge-only install, run those commands as brigadestack and tell your agents to do the same. Existing brigade installs need no change.
Works with
| Tool | Connection |
|---|
| Codex, Claude Code, Cursor, OpenCode | Workspace instructions, handoff inboxes, and tool and skill configuration. Hooks vary by tool. |
| T3 Code | Remote execution and session visibility. Full child-task governance is in progress. |
| Grok Bot | Role-scoped connector packs for bounded jobs. |
| OpenClaw, Hermes | Memory-owner workflows, writer handoffs, and memory care. |
| Dot | Memory, skills, Hub messaging, and authorized T3 fleet dispatch. |
| Duet (0.31 beta) | Memory, skills, Hub messaging, and read-only fleet views. |
Hosted connectors need explicit setup. See the technical guide, Dot guide, and Duet guide.
MCP servers
A memory server with bounded recall, work briefs, and handoff drafting, plus a read-only skills server:
brigadestack memory serve-mcp --stdio
brigadestack skills serve-mcp --stdio
write_handoff drafts a handoff and never edits a canonical card. Transport details
In the 0.31 beta
- Parallel workers. Dispatch ready tasks into separate worktrees, supervise them, and keep results for review.
- Memory reflection. Evidence-backed consolidation proposals for review. Nothing is rewritten automatically.
- Run patrol. Detects local lane states and offers explicit repair actions.
- In progress: fleet coordination, governed T3 delegation, cross-repository campaigns, and the agent mesh.
Reference
Commands · Execution model · Comparisons · Contributing · Issues · MIT license
The Python CLI may send a passive version request. Telemetry lists its fields, retention, and opt-outs.
The downloads badge sums brigade-cli on PyPI (excluding mirrors) and @escoffier-labs/brigade on npm over the same 30 completed UTC days. Downloads include automation and do not count unique people. The metric JSON records each source, the period, and the last successful refresh. npm platform companions are excluded. Daily refreshes retain the previous metric if a source fails validation or its latest day is more than three UTC days old.
Brigade Stack by Escoffier Labs. Project identity: GitHub escoffier-labs/brigade, website brigade.tools, PyPI brigadestack, npm brigadestack, command brigadestack. The brigade-cli and @escoffier-labs/brigade packages and brigade command remain supported.
