Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/coretracker/agentswarm/llms.txt

Use this file to discover all available pages before exploring further.

AgentSwarm is a Docker-based web application for running and managing AI coding work on real Git repositories. It gives developers and teams one place to create tasks, dispatch Codex or Claude agents into isolated workspaces, inspect streaming logs and diffs, review proposed changes before they land on a branch, and continue work in an interactive browser terminal — all without losing visibility into Git state, task history, or repository changes.

​
Key Features

Tasks

Create Build tasks that produce code changes or Ask tasks for read-only Q&A. Start from a blank prompt, a reusable snippet, a GitHub issue, or a pull request.

Agents

Run Codex (OpenAI) and Claude (Anthropic) agents in isolated Docker runtime containers. Stream logs, messages, and run status in real time from the task detail page.

Repositories

Connect any Git repository. Configure per-repository environment variables and secrets that are injected into agent runtimes at execution time.

Review Flow

Every build that produces changes creates a checkpoint — a reviewable change proposal. Apply, reject, or revert before pushing or merging to keep full control over what lands on a branch.

GitHub Integration

Attach a webhook to any repository and define automation rules that create tasks automatically from issue_opened, pull_request_opened, and comment or reaction triggers.

Sequences

Chain multiple prompts into a Sequence with two execution modes: automatically apply changes between steps, or pause for human approval before continuing.

Postflight

Define .agentswarm/postflight.yml in any repository to run automated checks (tests, screenshot captures, linters) after a successful build task and before the final checkpoint is created.

MCP Servers

Extend agent capabilities by connecting Model Context Protocol (MCP) servers over stdio or http transports, giving agents access to external tools and data sources.

​
Supported AI Providers

AgentSwarm supports two AI providers. Provider credentials are configured once in Settings and are never returned by the API after saving.

​
Codex (OpenAI)

Codex tasks run inside the agentswarm-agent-runtime-codex Docker image. Choose a model and reasoning effort level per task or set global defaults in Settings.
ModelValue
GPT-5.4gpt-5.4
o3o3
o4-minio4-mini
o3-minio3-mini
GPT-4.1gpt-4.1
GPT-4ogpt-4o
Effort levels: low · medium · high The default model is gpt-5.4. Effort maps to OpenAI’s native reasoning effort parameter.

​
Claude (Anthropic)

Claude tasks run inside the agentswarm-agent-runtime-claude Docker image. Claude supports an additional max effort level that leaves the thinking budget unconstrained.
ModelValue
Claude Opus 4claude-opus-4-5
Claude Sonnet 4.5claude-sonnet-4-5
Claude Sonnet 4claude-sonnet-4
Claude Haiku 3.5claude-haiku-3-5
Effort levels: low · medium · high · max The default model is claude-sonnet-4-5. Effort profiles map to thinking budgets on models that support extended thinking; max leaves the budget unset.

​
Architecture Overview

AgentSwarm is a monorepo with a clear separation between the web application, the API server, the shared type layer, and the agent runtimes. 
.
├── apps/
│   ├── server/           # Fastify API, orchestration, stores, routes, schedulers
│   └── web/              # Next.js web application
├── packages/
│   └── shared-types/     # Shared TypeScript types used by server and web
├── agent-runtime-codex/  # Automated Codex task runtime (Docker image)
├── agent-runtime-claude/ # Automated Claude task runtime (Docker image)
├── tools/                # Browser terminal and interactive runtime tooling
├── docker-compose.yml    # Local Docker stack
└── agentswarm.sh         # Main stack helper script
LayerTechnologyPurpose
apps/serverFastify (Node.js)REST API, task orchestration, queue workers, Git operations, webhook delivery
apps/webNext.jsBrowser dashboard, real-time task UI, interactive terminal
packages/shared-typesTypeScriptCanonical type definitions shared between server and web
agent-runtime-codexDockerIsolated execution environment for automated Codex tasks
agent-runtime-claudeDockerIsolated execution environment for automated Claude tasks
Durable data is stored in Postgres. Redis handles sessions, task queues, webhook jobs, and real-time pub/sub. Nginx proxies the web app and API on a single public port (default 3217).

​
Requirements

RequirementNotes
DockerRequired for the main app stack and agent runtime containers.
Docker Composedocker compose is preferred; docker-compose is also supported.
BashRequired by the helper and harness scripts.
Node.js 20+ and npmRequired for local development, checks, tests, and builds.
Python 3Required when installing local npm dependencies because native modules such as node-pty may build from source.

​
Next Steps

Quickstart

Clone the repo, run ./agentswarm.sh init, and create your first AI coding task in under 10 minutes.

Core Concepts

Learn about tasks, providers, checkpoints, sequences, snippets, and repository automation rules.

Build docs developers (and LLMs) love

Get started for freeTalk to us