Your AI Dev Team,
Running Locally
Spawn OpenAI Codex and Claude Code agents from a clean web UI. They plan features, write code, review pull requests, and push to GitHub — all inside isolated Docker containers on your own machine. No cloud. No data leaks. Just results.
Everything Your AI Agents Need to Ship
From a GitHub issue to a merged pull request — AgentSwarm handles the orchestration while you stay in control.
Full GitHub Integration
Connect your repositories and create tasks directly from GitHub issues or pull requests. Agents work on your real codebase, on your real branches.
Spawn Agents from a Web UI
No CLI needed. Open the web UI, pick a repo, choose a task type and complexity, select your agent — and hit go. Watch it work in real time.
Isolated Docker Environments
Every agent runs in its own sandboxed Docker container. Agents never touch your git credentials. The server reviews diffs and handles all pushes.
Live Terminal Streaming
See exactly what your agents are doing. Terminal output streams live to the web UI, with full history including plans, diffs, and what was integrated.
Flexible Branch Strategies
Create a dedicated feature branch per task or work on an existing branch. Every task gets its own worktree — and can even serve a live build for testing.
Integration Rules & MCP Support
Define per-repository or global integration rules that every agent must follow. Extend agents further with full MCP server support.
From Idea to Merged Code
AgentSwarm turns GitHub issues and blank prompts into working code — without leaving your browser.
Define the Task
Start from scratch, import a GitHub issue, or reference an open pull request. Set the source, task type, complexity level, and target branch. AgentSwarm keeps every option front and center so you never lose context.
Choose Your Agent & Complexity
Pick OpenAI Codex or Claude Code. Then dial in complexity — from a quick fix to a super-deep architectural pass. Each agent spins up in its own Docker container, scoped to your task.
Watch It Work
Terminal output streams live to the UI. Follow every command, edit, and decision the agent makes. Iterate as many times as you want — push updates or refine the task without starting over.
Review the Diff & Push
When the agent finishes, review the full diff right in the UI. See exactly what changed, approve it, and push. Build tasks are committed and pushed automatically. Plans stay as local markdown — yours to edit.
Best-in-Class Agents, Your Choice
AgentSwarm has full, native integration with both OpenAI Codex and Claude Code. Switch agents per task without changing your workflow.
OpenAI Codex
OpenAI's agentic coding model — fast, precise, and built for implementation. Ideal for feature builds, refactors, and targeted code changes across your codebase.
- Full native integration — not a wrapper
- Runs in an isolated Docker container
- Live terminal output streamed to your browser
- Respects your integration rules and MCP servers
Claude Code
Anthropic's frontier coding agent — exceptional at deep reasoning, planning, and complex multi-file tasks. A natural fit for architecture decisions, thorough code reviews, and intricate builds.
- Full native integration — not a wrapper
- Runs in an isolated Docker container
- Deep planning with local markdown output
- Respects your integration rules and MCP servers
Built for Every Phase of Development
Whether you're exploring, shipping, or reviewing — there's a task type for it. Mix and match across your workflow.
Plan
Let the agent think before it codes. Get a structured implementation plan as local markdown — yours to edit, approve, or hand back to the agent to build from.
Build
Hand a task to the agent and let it write, modify, and push code on a dedicated branch. Review the diff when it's done and ship with confidence.
Review & Ask
Get an AI code review on any pull request or diff. Ask natural-language questions about your codebase. Neither task touches your code — just insight, on demand.
Five Actions — Iterate Until It's Right
Every task supports unlimited iteration. Keep refining, adjusting, and pushing until the result is exactly what you wanted.
Generate a detailed implementation plan before writing a line of code
Implement features or fixes on a dedicated branch, ready to push
Refine the agent's output as many times as needed without starting over
Get a thorough review of any pull request or diff — no code changes
Ask anything about your codebase and get grounded, accurate answers
Your Code Never Leaves Your Machine
AgentSwarm is fully self-hosted. No cloud. No third-party data access. One docker compose up and you're running.
Web UI
A clean Next.js interface for managing repositories, spawning agents, tracking task history, and streaming live output. Open at localhost:3217.
Fastify + Redis Backend
Task orchestration, queue management, and credential storage all run server-side. Encrypted at rest. Never exposed via API.
Isolated Agent Containers
Each agent runs in its own short-lived Docker container. No git credentials inside. The server handles all commits and pushes after reviewing the diff.
Secure by Design
API keys and GitHub tokens are stored encrypted server-side, backed by httpOnly session cookies. Never visible in the UI or returned by the API.
Agents run in sandboxed containers without push access. No agent can push code without the server first computing and approving the diff.
Everything runs on your machine. Your code, your credentials, your data — completely offline-capable, with no telemetry or cloud dependency.
Up and Running
in Three Commands
No cloud account. No setup wizard. Clone the repo, run Docker Compose, and open the UI at localhost:3217. Your first agent task is minutes away.
$ git clone https://github.com/coretracker/agentswarm.git
$ cd agentswarm
$ docker compose up --build
→ AgentSwarm running at http://localhost:3217
→ Open Settings, add your GitHub token & API key
→ Add a repo, create a task, spawn an agent
Open source · MIT license