SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI codingbrigade

brigade

Local-first CLI (MIT) that keeps one reviewed source of the MCP servers, tools, skills, and memory your AI coding agents share, synced into each tool's native config (Claude Code, Codex, Cursor, VS Code, OpenCode, and more) with a review gate before every write and a rollback receipt. No daemon, no

Open website

About this tool

Brigade is a local-first command-line control plane for AI coding agents, developed by escoffier-labs and licensed under MIT. Its goal is to keep coding agents' work traceable across tools, repositories, and sessions, centered around tasks, receipts, shared memory, and synchronization tools. Regarding the release timeline, v0.27.0 is the released stable installation version (PyPI package name brigade-cli==0.27.0), and the main branch has entered the 0.28 testing line.

  1. Main capabilities Brigade provides the following capabilities:

  2. Ready work and claim safety: List unblocked tasks and use an atomic claim mechanism to prevent two agents from claiming the same task at the same time.

  3. Declared run limits, optional fleet hub, and registered cloud activity: Enforce declared wall-clock time and work dispatch limits, record named retry and rerouting decisions, report runs, repository claims, and external harness sessions through an operator-owned fleet hub, and reconcile registered cloud work with provider and GitHub status.

  4. Verification receipts: Checks executed through Brigade write the command, real exit code, and Git status. Note that not every Brigade command writes a receipt.

  5. Owner-mediated handoff and bounded recall: Handoffs are linted and routed by the memory owner. Safe and targeted notes may be archived automatically, while ambiguous or risky notes wait for review. Session-start recall is capped, and vault project, search, and propose are also bounded.

  6. Explicit dry-run-first harness projection: MCP servers, tools, and skills are previewed before Brigade writes to the harness.

  7. Outcome-based learning: Learning in Brigade is outcome-based skill scoring, promotion, and rollback from captured verification receipts, rather than autonomous reflective memory learning.

  8. How to use Installation: pipx install brigade-cli or uv tool install brigade-cli, then follow the installation guide, QUICKSTART.md, or hand the first 10 minutes and docs/agents-guide.md to your coding agent. Update channel: The stable channel is the released v0.27.0; the preview channelUse brigade update --channel beta to get the wheel package for 0.28.0.devYYYYMMDD. Execution model: Brigade is a local-file-first control plane; nothing runs unless a command or an operator-owned scheduler starts it. The optional serve commands (brigade center serve, brigade runs serve, brigade fleet serve) keep running in the foreground until manually stopped. Scheduling remains externalized, except for optional operator-owned, scoped care registration (brigade care install). Data storage: The work ledger is machine-local JSON, not a distributed task database. The optional fleet hub aggregates run events and arbitrates repository claims; its board can display external harness sessions, but local logs remain authoritative.

III. Applicable scenarios Suitable for developers or teams that need to uniformly manage MCP servers, tools, skills, and memory across multiple AI coding agent tools (Claude Code, Codex, Cursor, VS Code, OpenCode, etc.). Suitable for scenarios where work should be traceable, results verifiable, memory portable, and configuration synchronization controllable. Also suitable for scenarios that need to aggregate ready work across repositories and compose parallel safe waves (campaigns).

IV. Limitations Brigade is not a hosted memory service, an automatic release bot, or a complete autonomous fleet runtime. Unless the operator uses an explicit send operation, it will not push to GitHub, publish packages, or send chat messages. Ordinary runs have no hard run budget unless a budget is declared; even if declared, currently only wall-clock time and work dispatch limits are enforced, while model, tool, token, and cost dimensions remain observational unless the adapter owns the execution boundary. Cloud registry commands track work run by other providers. Center and fleet board are read-only. External activity observation is best-effort. Ambiguous or risky notes require review. Safe and targeted handoffs can be automatically archived under owner policy. Approved Grok Bot scout feeds are enqueued as labeled GitHub issues only after --apply.

V. Name conflict note This is the brigade brigade-cli from escoffier-labs/brigade, not the archived CNCF/Microsoft Kubernetes Brigade, Spinabot Brigade, nor the 2017 Python package that later became Nornir.

VI. Project identifiers GitHub: escoffier-labs/brigade; website: brigade.tools; PyPI: brigade-cli; command: brigade.

Related tools

Browse category