Multi-agent orchestration framework for AI coding assistants, featuring Lead/Worker architecture, DAG workflows, compounding memory, and multi-channel task intake.
Core Architecture#
Agent Swarm adopts a Lead/Worker orchestration pattern—the Lead Agent handles task intake and decomposition, while Worker Agents run in Docker containers providing isolated full development environments, supporting parallel processing across multiple repositories. Data storage is based on SQLite. The API server supports both HTTP (port 3013) and MCP protocols, providing an OpenAPI 3.1 specification.
Orchestration & Scheduling#
- DAG Workflow Engine: Directed acyclic graph-based automation with triggers, conditions, actions, retries, and structured I/O
- Human-in-the-Loop: Approval gates within workflows, supported via Dashboard UI and Slack notifications
- Scheduled & Periodic Tasks: cron / interval / delayed scheduling
Memory & Knowledge System#
- Compounding Memory: Agents extract learning outcomes after each session into shared memory, reusable for future tasks
- Persistent Identity: Each Agent has independent persona, expertise, and notes that evolve over time
- Skill System: Reusable programmatic knowledge that Agents can create, share, install, and publish
- Agent-fs: Persistent searchable filesystem
Multi-Channel Task Intake#
Supports Slack DM/@mention, GitHub App @mention/issue assign, GitLab webhooks, AgentMail (email → task), Linear bidirectional sync, and HTTP API task creation.
AI Backend Support#
- Claude Code (recommended harness)
- OpenAI Codex (API key / ChatGPT OAuth)
- pi-mono
- Gemini CLI (mentioned in README title, specific support status unconfirmed)
Observability & Operations#
- Real-time Dashboard: Monitor Agent status, task progress, inter-Agent conversations (local or hosted at app.agent-swarm.dev)
- MCP Server Management: Register, install, and manage MCP servers with scope cascade (agent → swarm → global)
- Context usage tracking, Sentry error tracking
Additional Features#
x402 USDC micropayments (details unconfirmed), Portless local dev domains, per-repo PR/Merge policy configuration, unified user identity resolution across platforms.
Quick Start#
Prerequisites: Docker, Claude Code OAuth token, Bun (>=1.0.26).
# Interactive wizard (recommended)
bunx @desplega.ai/agent-swarm onboard
# Manual deployment
git clone https://github.com/desplega-ai/agent-swarm.git
cd agent-swarm
cp .env.docker.example .env
# Edit .env — set API_KEY and CLAUDE_CODE_OAUTH_TOKEN
docker compose -f docker-compose.example.yml --env-file .env up -d
Unconfirmed Items#
- Gemini CLI specific support status
- x402 USDC micropayment implementation details
- Discord community invite link
- Worker Docker container base image and environment configuration details
- MCP protocol capability boundaries as Server/Client