Run a fleet of coding agents on your own cloud.

Every task gets its own agent and dev environment, on machines you own.
Agents can start helper agents, message them, and review what they send back.
Mix Claude Code, Codex, and other agents, and steer the whole fleet from your browser or phone.

Website • Docs • Deploy your own

_{A lead agent splits one request across helpers on Codex, Claude Code, and OpenCode. The real SAM interface with sample data.}

Built for more than a few agents

Running a few agents in terminal tabs works until your laptop becomes the bottleneck. They compete for ports, Docker, and memory. Your laptop has to stay awake. And each one stops to ask you something the moment you step away.

SAM moves the fleet off your laptop.

• A workspace for every task. On VMs in your cloud account, each agent gets its own dev container, built from your repo's .devcontainer by default, with Docker, git, and its own branch. Several agents share a VM when they fit, and VMs that SAM manages are removed after sitting idle for a while. For quick jobs, Instant sessions start in seconds on Cloudflare Containers, with no VM provider needed.
• Agents that run agents. An agent can start helpers, send them instructions, wait for them, and stop them. Ask a lead agent to split a job, hand out the pieces, and review what comes back.
• The agents you already use. Claude Code, Codex, Gemini CLI, Mistral Vibe, OpenCode, and Amp, side by side in one project. Save an agent, model, and effort level as a profile and pick one per task, so Claude Code can plan while Codex writes the code. Use API keys or your Claude and ChatGPT subscriptions.
• Everything visible, from anywhere. Every message and tool call streams into the project chat and stays there. The chat list shows which agents are working, finished, or waiting on you. Turn on push notifications to hear when a task finishes or an agent asks for your input.
• Work that starts without you. Schedules, GitHub events, and webhooks start agents: a nightly dependency sweep, a pass over every new issue, a first look at every production alert.
• Your accounts, your bill. SAM runs on your Cloudflare account. VMs run in your Hetzner, Scaleway, GCP, Vultr, UpCloud, DigitalOcean, or Infomaniak account, at their prices. Open source under AGPL-3.0.

Agents that run agents

Give one agent the whole job:

Add CSV export to the billing page. Break the work into the API, the UI, and the tests, run them in parallel, and review each branch before you tell me it is ready.

That is the conversation at the top of this page. The lead agent starts three helpers, each in its own workspace on its own branch. In this example, two helpers have finished, and the lead found a bug in the third and sent it back:

The lead does this with tools SAM gives its agents:

• dispatch_task starts a helper with a brief, a profile, and a branch.
• send_message_to_subtask sends a running helper new instructions.
• get_session_messages reads a helper's conversation.
• wait_for_subtasks pauses the lead until its helpers finish or a deadline passes. SAM wakes it up, so it never polls.
• stop_subtask and retry_subtask stop or retry a helper it started.
• start_workflow_run hands SAM a plan of steps that depend on each other.

Agents in the same project can message each other, so a helper can report back or ask the lead a question. Your deployment's settings cap how deep and how wide agents can fan out.

Plans with dependencies

For work in stages, an agent can hand SAM a workflow, such as "draft the release notes, have two agents review them at the same time, then apply both reviews". SAM starts each agent step once the steps it depends on finish, can wait for an event like a CI result between steps, and shows everyone in the project how the plan is going.

Your fleet from your phone

On your phone, see which agents are working, finished, or waiting on you, and answer the one that has a question. Here, the nightly dependency agent wants a decision before it takes on a major upgrade:

On machines you own

Every VM shows its provider, size, price, and load, and the agents running on it. Create VMs yourself, or let SAM create them as agents need them.

Teams and long-running work

Sessions sleep when they go idle and wake when you reply, with the conversation kept. Teammates can follow each other's agents, comment on an agent's reply, and share a chat for a set time. SAM tells every new session to load the project's knowledge and rules first. See what's in SAM.

How it works

Part	What it does	Built on
Control plane	API, auth, scheduling, chat history, agent-to-agent messaging	Cloudflare Workers, D1, Durable Objects, KV, R2
Nodes	VMs that run workspaces. Several agents can share one VM	Your cloud account, a Go agent on each VM, Docker
Workspaces	One dev container per task, with the agent, your repo, and Docker	Your repo's .devcontainer, or a fast prebuilt image
Instant	Containers for quick work when you don't need a VM	Cloudflare Containers

The control plane runs on Cloudflare Workers, so there is no control-plane VM to maintain. Read the architecture overview for the details.

Deploy your own

You need:

• A Cloudflare account on the Workers Paid plan ($5 a month minimum), with a domain on Cloudflare
• A GitHub App for sign-in and repository access
• A cloud provider account for VMs (optional if you only use Instant sessions)

SAM deploys from GitHub Actions in your fork. The self-hosting guide walks through every step.

Develop

pnpm install && pnpm build && pnpm test

See the local development guide to run SAM on your machine.

Learn more

• Docs: agents, workflows, automation, and teams
• Self-hosting guide: deploy your own instance
• Security model: how credentials are encrypted and isolated
• How SAM compares: with vendor cloud agents and self-hosted dev environments
• Blog: what we are building and why

License

AGPL-3.0