Brigade is a self-hosted multi-agent system: a crew of AI agents arranged on an org chart that share one long-term memory, called Tideline, so facts one agent learns are available to the others. Agents delegate to each other, switch models mid-conversation without losing context, and act in external apps through connectors. It runs as a local CLI on anything from a Raspberry Pi to a server, with no account and no telemetry, and keys and data stay on the owner's machine.
One binary acts either as a gateway, the long-lived process that holds all state and runs agents, or as a thin client attached to it over WebSocket. The terminal chat interface, messaging channels, and MCP clients all route through the same pipeline, which resolves each inbound message to an agent and session so that contexts on different channels stay separate. Each turn assembles the persona, skills, and org context, streams the model, runs tool calls through approval and ownership guards, and persists the transcript. A heartbeat and an out-of-process supervisor restart the gateway if it stalls.
Every agent has its own workspace, persona, credentials, memory, sessions, and approvals. An agent-to-agent policy derived from the org chart governs which agents may message which. State lives in the filesystem by default, with an optional bundled Convex backend that can be self-hosted. Models can be reached with API keys, with a Claude, ChatGPT, or GitHub Copilot subscription, or by reusing an existing Claude Code or Codex login, and the Claude Code binary can serve as the engine with Brigade's own guarded tools supplied to it over MCP.
Features
- Tideline memory: origin-scoped facts with decay, write-time deduplication, a provenance write-gate, hybrid keyword and vector recall, and a nightly consolidation pass
- Org chart: members, departments, and reporting lines that determine which agents can talk to which
- Sub-agents: single or parallel fan-out delegation with depth and child-count caps and an abort cascade
- Team Mode: durable rooms and multi-agent task graphs with dependency joins, retries, budgets, approvals, handoffs, and restart recovery, exposed as a headless backend
- Bring any model: Anthropic, OpenAI, Gemini, OpenRouter, Groq, Cerebras, xAI, DeepSeek, Mistral, Ollama, and custom OpenAI-compatible endpoints, switchable mid-conversation
- Channels: WhatsApp, Telegram, Slack, Discord, iMessage, and BlueBubbles through a typed adapter contract
- App connectors: more than 1,000 apps such as Gmail, GitHub, Notion, and Linear through a built-in Composio tool with managed OAuth
- Documents and media: reads PDFs, Office files, images, audio, and video, and creates or edits Word, Excel, PowerPoint, and PDF files
- Skills: folder-based skills discovered automatically, with 56 bundled
- Scheduling: cron expressions, intervals, or one-shot jobs that run agent turns, system events, or shell scripts and survive restarts
- MCP: exposes long-term memory to any MCP client and connects external MCP servers
- Extension SDK: typed slots for tools, hooks, commands, model providers, channels, web search, and gateway methods
