AI workers on your own machine

Give your AI a team ofworkers.

Name: Swarm
Author: SwarmAgent

Swarm runs on your own machine. Your AI (Claude Code, Codex or ChatGPT) gives it work: code, apps, designs, research, images and video, on a schedule if you like. You see what happened and what it cost.

Paste this into your AI

Read swarmagent.dev/llms.txt and set up Swarm on this machine for me.

Or run it yourselfcurl -fsSL https://swarmagent.dev/install | sudo bash

Works with Claude Code, Codex and ChatGPT

Ubuntu 24.04 (x86-64) on your PC, a server or a VM.

What you can ask

Say it the way you would say it to a person.

Each of these is a real request you can type to your AI today. Your AI hands the work to Swarm and tells you how it went.

Fix the failing tests in my repo.
The coder works on its own branch in a sandbox. Your AI reviews the fix and merges it.
Build me a small booking app and show me it works.
The coder builds it, runs it and reports what it checked. Your AI looks at the result before merging.
Every morning at 8, check my site and tell me if anything broke.
Your AI puts the job on a schedule. Every run is a job it can read back to you.
Mock up a landing page for my product.
The designer drafts it on its own branch, so you see it before anything changes.
Make a 10-second product video.
Swarm asks a video model you connected and saves the file for your AI to hand you.
Use a faster model for the coder.
Your AI switches the coder’s model. The other workers stay on their recommended models.

What you get

A team, a schedule and a clear record.

Everything runs on your machine. Your AI does the asking, Swarm does the work, and you can always see what happened.

Workers and jobs
A coder, a designer and a finder on recommended models, plus your own workers with standing instructions. Each job runs on its own branch; your AI reviews it and merges it.
Automations
Put any job on a schedule, with cron or every N minutes. Each run is a job you can read.
Media
Images, video and sound from the models you connect: your ChatGPT plan’s image model, Google or OpenRouter.
Monitoring
Ask “is everything OK?” and get the status, what happened since last time, what failed and what it cost. Set spending limits.
Networking
Private by default: Swarm listens on 127.0.0.1 only. Reach it from your other devices over Tailscale, from ChatGPT through OpenAI’s tunnel, and from claude.ai or ChatGPT on the web through your own Cloudflare relay. No public ports.
Safety
Every job runs in a gVisor sandbox with no access to your machine, network or keys. Secrets reach a job only through Swarm’s gateway, and agents see a stand-in. Keys never go through chat.

Works with

The AI you already use.

Keep talking to the AI app you like. Swarm’s workers run on the models you connect.

AI apps

Claude CodeLive
Connects over local MCP with the key the installer leaves for it.
CodexLive
Connects over local MCP with the same key.
ChatGPTPreview
Connects through OpenAI’s tunnel, with no public ports.
claude.ai and ChatGPT on the webPreview
Connect through your own Cloudflare relay. Your machine dials out; nothing listens on it.
Your other devicesPreview
Reach Swarm over Tailscale from your laptop or phone.

And anything else that speaks MCP.

Models

Your ChatGPT plan
OpenAI
Anthropic
Google
OpenRouter
Fireworks

Start with your ChatGPT plan or any one API key. Swarm gives each worker a recommended model, and your AI can change it when you ask.

How it works

Three steps, all in your AI’s chat.

Paste the prompt into your AI
Paste “Read swarmagent.dev/llms.txt and set up Swarm on this machine for me.” into Claude Code, Codex or ChatGPT.
Your AI sets it up
It installs Swarm, connects your ChatGPT plan or an API key, and asks which folders your work is in.
Ask for work
Your AI gives the work to Swarm’s workers, checks the result and merges it.

Build on it

Everything is a tool your code can call.

Everything Swarm does is an MCP tool over HTTP. Anything your AI can do with it, your own program can do too: give jobs, wait for them, read results and costs. Build your own app or orchestrator on top.

Swarm is open source under Apache 2.0.

FAQ

Common questions.

What is Swarm?

Swarm gives your AI a team of workers on your own machine. Your AI (Claude Code, Codex or ChatGPT) gives them jobs: code, apps, designs, research, images and video, now or on a schedule. Then it reads what they delivered and what it cost, and merges the result. Swarm is open source under Apache 2.0.

What does it run on?

Ubuntu 24.04 on x86-64: your own PC, a server or a VM. On a Mac, use an Ubuntu 24.04 VM.

Do I need an API key?

No. Your ChatGPT plan works. You can also add API keys for OpenAI, Anthropic, Google, OpenRouter or Fireworks. You add keys yourself on the machine; they never go through chat.

Is it safe?

Every job runs in a gVisor sandbox with no access to your machine, your network or your keys. Secrets reach a job only through Swarm’s gateway, and agents only see a stand-in. Each job works on its own branch, and nothing reaches your folder until your AI merges it. Swarm listens on 127.0.0.1 only and opens no public ports.

Can I use it from my phone or laptop?

Yes, over Tailscale (in preview). Your AI on your other devices reaches Swarm over your private tailnet, with no public ports.

Can I build my own app on it?

Yes. Everything Swarm does is an MCP tool you can call over HTTP, so a program can do anything your AI can: give jobs, wait for them, read results and costs. Build your own app or orchestrator on top.

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.