swarmagent.dev
GitHub
DocsModelsResourcesPricingNewsContactGitHub

Headless UX for any idea

Build yourbot.

Describe the idea. The AI builds it for you, headless: the logic and agents run on your machine, and you drive it from Claude, ChatGPT, your phone, or your own UI. You approve what ships.

Plugs into
  • Claude · preview
  • ChatGPT · preview
  • Tailscale
  • Local
  • Swarm Control relay · preview
  • Your own UI · preview
  • Instant deploy · soon
Build me a Telegram bot that posts a summary of yesterday’s commits every morning.
  1. bot
  2. app
  3. agent
  4. tool
  5. studio
curl -fsSL https://swarmagent.dev/install | sh
Learn More
QuickstartView GitHub

Linux and WSL today. Native macOS in progress.

Headless by design

One build. Every interface.

What Swarm builds is not stuck in one window. The logic, agents, and state live in one always-on runtime that you own, and any surface can plug into it.

Swarm Control · Preview

Run your swarm from Claude and ChatGPT.

Swarm Control gives AI clients one OAuth-protected MCP URL for your machines. Deploy the open-source relay to your own Cloudflare account, add it as a custom connector, and start sessions, check progress, and approve work from chat.

  • Your machines dial out. Nothing listens on them.
  • Off by default. You turn it on per machine.
  • New AI clients are approved on your own machine, not on a web page.
  • Each machine sets its own limits for writing, approving, and managing.
  1. Claude · ChatGPTHTTPS + OAuth
  2. Your relay /mcpyour Cloudflare account
  3. Your machinesoutbound connection only
# on each machine
swarmctl remote init --relay "$RELAY_URL" --name "Box A"
swarmctl remote enable

# when Claude shows a code
swarmctl remote approve ABCD-EFGH

Your own UI · Developer preview

Put any front end on what the AI built.

The TypeScript SDK drives Swarm from your own backend: create a session, send a message, stream the reply, and show permission requests in your own UI. Keep daemon credentials on the server, never in the browser. It is an early preview, so expect the API to change.

npm install @swarm-agent/sdk
import { SwarmClient } from '@swarm-agent/sdk';

// In your own backend, next to Swarm.
const client = new SwarmClient({
  socketPath: process.env.SWARM_SOCKET_PATH,
});
const session = await client.chat.createSession({
  title: 'My bot',
  workspace_path: '/path/to/project',
});

const { reply } = await client.chat.run(session.id, {
  message: 'Add a /summary command to the bot.',
});

How it works

You bring the idea. The AI builds it.

No stack to pick, no boilerplate to write. Swarm does the building and keeps you in the loop at the moments that matter.

  1. Describe the idea

    Type it the way you would tell a friend. No settings, no prompt engineering, no config files to learn first.

  2. The AI builds it

    A lead agent turns your idea into a plan and hands pieces to specialists. Each one builds and tests in its own isolated copy of the project, so nothing collides.

  3. Plug in any interface

    What you get is headless: use it from the terminal, the browser, your phone over Tailscale, or your own UI. Approve changes before they ship.

swarm · telegram-bot
  • swarmLeadPlanned 3 tasks for “Telegram commit-summary bot”
  • finderResearchRead the repo layout and existing config
  • coderBuildWrote bot.py and tests in an isolated worktree
  • coderCheckRunning tests…
  • youApproveReview the commit when it is ready
Illustration of a run. Real runs show live progress like this in the terminal and browser.

What will you build?

Start with a sentence.

Each of these is a real request you could type today. Swarm picks the right agents for the job.

Provider compatibility

Use the provider you prefer.

OpenAI

OpenAI API and Codex sign-in paths

Anthropic

Claude model access

Google

Gemini model access

Fireworks

Hosted model inference

OpenRouter

Multi-provider model routing

Coming soon

More provider integrations are on the way.

Trust Architecture

Your work belongs to you.

Swarm keeps control in your hands by combining permission boundaries, inspectable source, and durable session visibility in one unified system.

PERMISSIONS

Security

Tool, shell, and workspace access can require explicit approval before work continues.

  • Explicit permission prompts
  • Scoped workspace access
  • Configurable provider auth
INSPECTABLE

Open Source

The Swarm codebase is public, so you can inspect how the orchestrator and runtime work.

  • Public source repository
  • Reviewable runtime paths
  • Open development history
DURABLE STATE

Observability

Durable sessions, checkpoints, and live updates keep work visible as it progresses.

  • Live session updates
  • Durable session history
  • Checkpoint status and review

Agent-readable

Let your AI set it up.

This site is written for people and for the agents working for them. Paste this into the assistant you already use and it can find everything it needs.

Read https://swarmagent.dev/llms.txt, then help me install Swarm and have it build my idea as a headless app.

Index for agents
https://swarmagent.dev/llms.txtShort, link-rich map of the docs.
Full reference
https://swarmagent.dev/llms-full.txtAll core docs in one file for a single context window.
Install
curl -fsSL https://swarmagent.dev/install | shInteractive. Shows an install plan before changing anything.
Unattended install
curl -fsSL https://swarmagent.dev/install | sh -s -- --yes --serviceNo prompt; installs and starts the systemd service.
Headless SDK
npm install @swarm-agent/sdkDeveloper preview. Drive sessions from code.

FAQ

Common questions.

What is Swarm?

Swarm turns an idea into a working headless app, and the AI does the building. You describe what you want in plain language; a lead agent plans it and hands the work to specialist agents that build and test it in isolated Git worktrees. You review and approve the result. Swarm is open source and runs on your own Linux machine or server with the AI provider you choose.

What does “headless” mean here?

What you build is not locked to one screen. Its logic, agents, and state run in Swarm on your machine, and interfaces plug into it: the terminal, the browser view, your phone over Tailscale, your own front end through the TypeScript SDK, or Claude and ChatGPT through the Swarm Control relay (in preview). One-step deploys are coming soon.

Do I need to know how to code?

No. You describe the result you want in everyday words and review what comes back. Swarm builds inside a Git repository, and its permission rules can pause for your approval before it runs commands or changes files, so you stay in control.

Which platforms can run Swarm?

The current installer supports Linux x86-64, including WSL. On a Mac, run Swarm inside a Linux x86-64 VM or suitable Linux container; the Linux guest, not macOS, is the supported environment. A systemd host can manage the service, or you can install the files only and use your own supervisor.

Can I put my own interface on top?

Yes. The TypeScript SDK, @swarm-agent/sdk, is available on npm as an early developer preview. Run it in your own backend to create sessions, send messages, stream replies, and handle permission requests in your own UI. Keep daemon credentials on the server, never in browser code, and expect the API to change before a stable release.

Can my AI assistant set Swarm up for me?

Yes. Point it at https://swarmagent.dev/llms.txt for a compact, machine-readable index, or https://swarmagent.dev/llms-full.txt for the full reference. The installer also accepts --yes and --service for unattended installs.

Can I use Swarm from Claude or ChatGPT?

Yes, in preview. Swarm Control is an MCP endpoint built into Swarm. For Claude and ChatGPT on the web, deploy the open-source Swarm Control relay to your own Cloudflare account and add its /mcp URL as a custom connector. Your machines dial out to the relay, nothing listens on them, remote access stays off until you enable it, and every new AI client must be approved from one of your own machines with swarmctl remote approve.

How do I reach Swarm from my phone or another computer?

Use Tailscale Serve, which Swarm can verify before allowing access to its browser interface, then install the PWA from your mobile browser. To work through Claude or ChatGPT instead, use the Swarm Control relay. Do not expose Swarm’s local ports directly to the internet.

Can Swarm host this for me?

We are opening a waitlist for hosted Swarm plans from $19.99 per month: Your Personal AI Studio with a rolling temporary-compute allowance. Dedicated servers, private setups, and persistent Always-On Nodes remain available for interest by request only. There is no self-service purchase or automatic provisioning yet.

What will you build first?

Install Swarm, connect the AI provider you already use, and describe your idea. The AI builds it headless, so you can use it from your terminal, your browser, your phone, or your own UI. You stay in charge.

Read the quickstartView GitHub