Agent Swarm

What Agent Swarm does

Agent Swarm provides an API and MCP interface for coordinating multiple AI agents around software and operational work. A lead agent receives goals and delegates them to worker agents, which run in isolated Docker containers with development environments. The system keeps shared memory and identity across sessions, allowing later work to reuse accumulated context and agent-specific instructions.

Tasks can enter through Slack, repositories, issue trackers, email, the API, or the CLI. Completed work can produce structured JSON results, pull requests, Slack replies, or email replies. Workflows, scripts, schedules, apps, review gates, and extensions support recurring or controlled automation. Realtime rooms provide shared page state, presence, and live channels.

How it works

The lead agent plans and distributes work to one or more workers. Workers read shared context from persistent memory and write back learnings. The deployment model uses Docker containers for worker isolation, while the self-hosted installation stores orchestration data locally with SQLite.

The MCP-facing task flow includes a send-task operation for callers that need schema-validated results. Deferred tasks can wait until all or any watched tasks reach an outcome, or until a deadline occurs. This makes the server suitable for workflows where one agent’s output controls the next step rather than requiring every task to run immediately.

Agent Swarm supports several agent harnesses and model providers, including Claude Code, Codex, pi, opencode, Cursor, Amp, Grok, Devin, and ACP agents. Integrations cover Slack, GitHub, GitLab, Azure DevOps, Linear, Jira, AgentMail, WhatsApp through Kapso, Telegram, Teams, Discord through Chat SDK, Composio, Sentry, and Serply search.

Setup and configuration

The reference deployment uses Docker Compose. A one-shot installer checks for Docker and Compose v2, installs Bun if needed, runs onboarding, and prints a dashboard setup link. The installer defaults to the single-agent solo preset; --preset full enables the official full template set.

A manual setup clones the repository, copies .env.docker.example to .env, sets API_KEY and credentials for the selected HARNESS_PROVIDER, creates an encryption_key file, and starts the example Compose configuration. The local API is available at http://localhost:3013, with documentation at /docs and the OpenAPI description at /openapi.json. Kubernetes deployment is also supported through the repository’s OCI Helm chart.

Agent identity variables are optional. Containers generate a UUID on first startup and retain it in their personal volume; pinning an *_AGENT_ID value preserves identity when volumes are removed or an identity must be explicitly fixed.

Tools and capabilities

• Submit tasks and request schema-validated JSON results.
• Defer tasks until watched tasks finish or a deadline arrives.
• Coordinate lead and worker agents in isolated Docker environments.
• Preserve memory and identity between sessions.
• Run workflows, schedules, scripts, and apps for recurring work.
• Attach files and display inline tool-result images in task records.
• Use trusted TypeScript extensions activated by an authorized lead, operator, or dashboard user.
• Connect work intake and output to supported communication, repository, issue-tracking, and monitoring integrations.

Limitations and notes

The self-hosted software is MIT-licensed, while the website and cloud service have separate privacy and terms links. Docker and Compose v2 are required for the documented Compose installation. The full preset deploys one lead and multiple workers, so resource requirements depend on the selected agent set. AgentMail archives verified inbound deliveries for 30 days, including messages excluded from task routing.