For AI agents

eutonomous mcp offers your workspaces to an AI agent as tools, over the Model Context Protocol. The agent can create a workspace, run commands and Python in it, read and write files, deploy a project and delete the workspace when it is done.

Set it up

Install the command and sign in: uv tool install eutonomous, then eutonomous login.
Add the server to your agent’s MCP configuration:

{
  "mcpServers": {
    "eutonomous": { "command": "eutonomous", "args": ["mcp"] }
  }
}

The agent’s client starts the command itself; there is nothing to keep running.

What the agent may do

The agent acts with a key, and can do what that key may do and nothing more. By default that is the key from eutonomous login. To give the agent less, create a key on the platform’s CLI keys page and pass it in the configuration:

{
  "mcpServers": {
    "eutonomous": {
      "command": "eutonomous",
      "args": ["mcp"],
      "env": { "EUTONOMOUS_API_KEY": "eut_…" }
    }
  }
}

Two set-ups that work well:

Its own sandboxes only. A key limited to one environment you do not mind, with deploy and create. The agent can create workspaces and use the ones it created, and cannot touch your others.
One existing environment. A key limited to that environment with deploy only. The agent can run commands and deploy there, and cannot create or delete anything.

Every key expires, and what the agent does with it is in the platform’s audit log. See Keys and CI.

Tools

Tool	What it does
`list_workspaces`	the environments the key can use, with state and URL
`create_workspace`	a new workspace; the first command waits until it is set up
`delete_workspace`	delete a workspace and its files; needs `confirm: true`
`sleep_workspace`	put a workspace to sleep; the next command wakes it
`snapshot_workspace`, `list_snapshots`, `restore_snapshot`	keep a workspace’s files and go back to them
`exec`	run a shell command, returns output and exit code
`run_python`	run a piece of Python, returns what it printed
`read_file`, `write_file`, `list_files`	files in the workspace
`deploy`	upload a project directory from your computer and serve it
`logs`	the deployed app’s output

A sleeping workspace wakes on the first command, in a few seconds. Long command output is shortened in the middle before it goes back to the agent.