Brigade runs a crew of AI agents on hardware you control, coordinated through a defined org chart and sharing one long-term memory called Tideline, so a fact learned by one agent is available to the rest. It installs as a single binary that runs either as a gateway — the long-lived process that holds all state and runs agents — or as a thin client that attaches to that gateway over WebSocket, so you can disconnect and reconnect while agents keep working.
Agents are spawned with their own personas, credentials, and workspaces, delegate to child agents singly or as a parallel fan-out, and can change model provider mid-conversation while keeping the transcript. Supported providers include Anthropic, OpenAI, Gemini, OpenRouter, Groq, Cerebras, xAI, DeepSeek, Mistral, a local Ollama, and any OpenAI-compatible endpoint; an existing Claude Code or Codex CLI login can be reused in place of an API key.
State lives under a single ~/.brigade/ directory, on the filesystem by default or in a self-hosted Convex backend with AES-256-GCM at-rest encryption. Privileged tool calls pass a per-call ownership gate, the shell tool runs behind a per-agent approval allowlist, and the gateway is localhost-only and unauthenticated by default, with optional per-device access tokens.
Features
- Long-term memory: facts persist across sessions with origin scoping, three-tier decay, hybrid keyword and vector recall, a typed link graph, and a nightly consolidation pass
- Skills: a folder with a SKILL.md is discovered on the next turn, with name and description prompt-resident and the body read on demand; 56 skills ship bundled
- Sub-agents: single and parallel delegation to child agents, depth- and count-capped, with an abort cascade and a completion bridge back to the parent
- Org chart: members, departments, and reporting lines define an agent-to-agent policy governing which agents may message which
- Channels: WhatsApp, Telegram, Slack, Discord, iMessage, and BlueBubbles adapters share one inbound pipeline for access control, dedupe, and approval routing
- Scheduling: cron expressions with timezones, fixed intervals, and one-shot timestamps; payloads can be a system event, an isolated agent turn, or a shell script, and jobs survive restarts
- App connectors: an owner-only Composio tool reaches 1,000+ connected apps such as Gmail, Slack, GitHub, Notion, Calendar, and Linear with managed OAuth
- Documents and media: reads PDFs, Word, Excel, PowerPoint, images, audio, and video, and creates or edits Word, Excel, PowerPoint, and PDF files in place
- MCP: exposes long-term memory to any MCP client over stdio, and connects external MCP servers in
- Model switching: the full transcript carries onto a new model, thinking level is re-anchored to what the target supports, and provider-specific reasoning blocks are sanitised
- Autonomous loops: a loop runner drives an agent under token, iteration, and time budgets with no-progress and repetition detection, stopping on independent done-checks
- Storage modes: filesystem by default, or a self-hosted Convex backend with at-rest encryption
