agent-swarm.dev: AI Agent Operating System for Teams
desplega-ai/agent-swarm
agent-swarm.dev is a self-hosted platform for running teams of Claude Code, Codex, and other AI workers under one lead agent that plans, delegates, and remembers across sessions. Reach for it if you already run Docker and want agents that write PRs, Slack replies, and reports from a shared memory store. Skip it if you just want a single coding assistant — the multi-agent orchestration is overkill for that.
What is agent-swarm.dev?
agent-swarm.dev is a self-hosted system that turns Claude Code, Codex, and other AI coding tools into a coordinated team: one lead agent splits an incoming goal into tasks and hands them to worker agents, each sandboxed in its own Docker container. Workers ship pull requests, Slack replies, or generated pages, then write what they learned into a shared memory store so future tasks start with more context.
Why become AI-Native with agent-swarm.dev?
Most companies bolt AI onto existing tools — a Slack bot here, a Copilot plugin there — and each integration starts from zero context every time. agent-swarm.dev's own framing is "AI First" versus "AI Native": task delegation and learnings stay siloed per tool instead of compounding. The README's fix is an LLM orchestration platform built around one lead agent and a shared memory store that every department's workers read from and write to, so a support agent's discovery becomes something the engineering worker already knows next week.
Who is agent-swarm.dev for?
agent-swarm.dev fits engineering-led teams that already run Docker and want self-hosted AI agents routing work from Slack, GitHub, or Linear instead of a single chat window. As an agentic AI platform, it suits companies willing to operate their own containers and credentials for Claude Code, Codex, or another supported harness, since nothing here is a hosted SaaS with a login page. It's a poor fit for a solo developer who just wants inline code completion, or a team that isn't ready to hand a worker agent write access to a repo or ticket queue.
Key Features of agent-swarm.dev
- ✓Lead/worker split running in Docker — the core of agent-swarm.dev's multi-agent framework: the lead agent plans and delegates, while workers execute in isolated containers with their own dev environment (git, Node.js, Python).
- ✓Persistent memory with hybrid recall that blends vector and full-text search, follows links between related memories, and lets an agent correct a memory without losing its ID or history.
- ✓A DAG-based workflow engine with human-in-the-loop approval gates, retries, and structured input/output for steps that shouldn't run unattended.
- ✓Cron-based scheduling that can trigger agent tasks, workflows, or catalog scripts on a recurring basis.
- ✓Mid-run task steering involves adding context at the next turn boundary, or alternatively, an immediate interrupt can be triggered when the harness provides support.
- ✓Harness- and LLM-agnostic execution: Claude Code, Claude Bridge, OpenAI Codex, pi-mono (Anthropic, OpenRouter, or Bedrock), Devin, Claude Managed Agents, raw LLMs, or opencode, with per-task modelTier intent (smol, regular, smart, ultra).
- ✓Observability via OpenTelemetry traces plus OTLP cost/token metrics for dashboarding session cost in SigNoz, Datadog, or Tempo.
- ✓DB-backed pages and a Redis-like KV store so agents can publish reports or share scoped state without standing up a separate database.
Practical Use Cases for agent-swarm.dev
- •Feature development: a Slack request becomes a Linear ticket and a GitHub pull request without a human relaying context between tools.
- •Workflow automation for code health: Datadog, New Relic, or Sentry alerts kick off a worker that investigates and proposes a fix on a schedule.
- •Proactive customer support: an agent watches specific accounts and prepares scheduled status reports instead of waiting for a ticket.
- •Reporting from a data warehouse: agents query fresh data and answer standing questions a BI dashboard used to handle.
- •Release documentation: workers update docs and, per the README, can pair with tools like Remotion or browser-use to generate release videos.
Getting Started with agent-swarm.dev
The fastest path is the onboarding wizard, which collects credentials, picks presets, and writes a working docker-compose.yml for you: ```bash bunx @desplega.ai/agent-swarm onboard ``` For manual setup, clone the repo, copy the example env file, fill in API_KEY plus a harness credential (for example CLAUDE_CODE_OAUTH_TOKEN), and bring it up with Docker Compose: ```bash git clone https://github.com/desplega-ai/agent-swarm.git cd agent-swarm cp .env.docker.example .env docker compose -f docker-compose.example.yml --env-file .env up -d ``` Docker plus at least one supported harness credential is a prerequisite — the default quick start assumes a Claude Code OAuth token from `claude setup-token`. Once running, the API listens on port 3013, with interactive docs at /docs and an OpenAPI 3.1 spec at /openapi.json.
Strengths
- ✓MIT-licensed and self-hosted, so the memory store, task history, and credentials stay on your own infrastructure.
- ✓One lead agent instead of a pile of point integrations — Slack, GitHub, GitLab, Linear, Jira, and email all feed the same task queue.
- ✓The memory system is more than a vector store: it links related memories and lets agents amend an existing memory instead of only appending new ones.
- ✓Not locked to one model vendor — Claude Code, Codex, Devin, and raw LLMs via pi-mono all plug into the same lead/worker structure.
- ✓Ships real production concerns out of the box: OpenTelemetry traces, cost/token metrics, and a systemd-based deployment guide, not just a demo docker-compose file.
Considerations for agent-swarm.dev
- △This is infrastructure you run and patch yourself — there's no hosted SaaS tier, so you're responsible for the Docker host, the SQLite database, and every harness credential.
- △Handing worker agents write access to a repo, a ticket queue, or Slack means a bad task can ship a bad PR or send a bad message; the human-in-the-loop gates only cover the steps you configure as approval points.
- △The README says the repo "evolves every single day," which is a fast-moving codebase to build production automation on top of — expect breaking changes between releases.
- △Harness credentials (Claude Code OAuth token, Codex, Bedrock, etc.) are a separate cost and setup step per provider you want to use.
Alternatives to agent-swarm.dev
Frequently Asked Questions
agent-swarm.dev is an open-source, self-hosted system that runs a lead AI agent alongside worker agents in Docker containers, coordinating tasks from Slack, GitHub, GitLab, Linear, Jira, email, or an API, and sharing memory across sessions.
agent-swarm.dev connects to Slack, GitHub, GitLab, AgentMail, Kapso for WhatsApp, Composio, Linear, Jira Cloud, Sentry, and Devin per the integrations table in the README, plus a general HTTP API and CLI for anything not covered.
agent-swarm.dev works with Claude Code, Claude Bridge, OpenAI Codex, pi-mono (which reaches Anthropic, OpenRouter, or Amazon Bedrock), Devin, Claude Managed Agents, raw LLMs, and opencode, with per-task model tiers from smol to ultra.
agent-swarm.dev is released under the MIT license, so you can self-host, modify, and redeploy it without a commercial license fee.
agent-swarm.dev ships a dedicated deployment guide and a DEPLOYMENT.md covering Docker Compose with multiple workers, systemd for the API, and graceful shutdown, so production deployment is documented rather than assumed.
agent-swarm.dev stores memory with hybrid vector and full-text recall, links related memories together, and enables workers to modify an existing memory, ensuring its ID or history remains intact; identity persists through files like SOUL and CLAUDE.md, per the architecture docs.