One tool. The whole editor.

One tool, not 347

A narrow tool does one thing per round trip, and its schema rides along in every request. Satura Space registers one. The agent writes the body of an async function, and every call in it settles when the work is really done: Play entered, compile finished, video saved. One script compiles, plays, drives the game, films it and opens a report.

• The agent presses Play and reads the console itself, so you stop pasting errors back.
• One reply per script instead of one round trip per step, and no 347-tool schema in every request.
• unity.exec reaches every Editor API and package you have, with no recompile or domain reload.
• A modal dialog throws DialogBlocked in about a second instead of hanging, and unity.answer presses its button.
• The same script opens a Tuner when a value needs your eye.

It's written into .mcp.json and the Claude Code, Cursor and Codex configs for you, with no API key. Other CLI agents still work in a normal shell.

What a script looks like

Each line waits for the one before it. Failures are exceptions the agent can branch on: CompileError, HandleError, DialogBlocked, UnityTimeout.

await unity.refresh()
await unity.play('Assets/Scenes/Pond.unity')
await handles.call('PrepareStageForFishTest')
const bite = await handles.call('WaitForBite', 10)
await unity.stop()
return bite.value

When a value needs your eye, the agent hands it over and ends its turn:

return await tuner.open({
  name: 'arena-lights',
  note: 'Balance the arena lights, then copy your changes to me',
  objects: [{ path: '**', components: [
    { type: 'Light', fields: ['m_Intensity', 'm_Color'] }
  ] }]
})

The reply is the return value plus each step with its time, the failed one marked. Scripts get no file access and run up to 300 seconds by default.