Agent Swarm 🤖 Agent Open source

What it does

Agent Swarm is an open-source "operating system" for running AI coding agents as a persistent, multi-agent workforce rather than one-off terminal sessions. A lead agent receives work from Slack, a repository, an issue tracker, email or an API, then delegates it to worker agents that each run in an isolated Docker container with its own development environment.

What is inside

• A lead/worker delegation model with task scheduling and review gates
• Persistent memory per agent, with vector search and an identity that survives across sessions
• A dashboard with file attachments and inline tool-result images
• Deferred tasks that fire on dependencies or deadlines
• Real-time collaboration rooms with shared page state
• Structured output validation for workflows, scripts and applications

Works with

The README lists Claude Code, Codex, Cursor, Gemini CLI, Devin, pi, OpenCode and Kimi as supported coding agents that can act as workers, alongside its own containerized workers.

How to install

Add it to an existing harness:

npx skills add desplega-ai/agent-swarm

Or run the full stack with Docker Compose:

git clone https://github.com/desplega-ai/agent-swarm.git && cd agent-swarm
cp .env.docker.example .env
openssl rand -base64 32 > encryption_key
chmod 600 .env encryption_key
docker compose -f docker-compose.example.yml --env-file .env up -d

Maintenance and safety

MIT licensed (desplega.sh, 2025-2026), with 832 stars, 105 forks and a very active commit history of over 2,300 commits as of late September 2026. Self-hosting it means running Docker containers and holding an encryption key that protects stored credentials and memory, so treat the deployment like any other backend infrastructure: secure the .env and key files and scope container permissions carefully.

Who should use it

Teams or solo builders who have outgrown single-session coding agents and want a persistent, multi-agent system that can pick up work from Slack or an issue tracker, delegate it to isolated workers, and report back, rather than one terminal loop per task.