# Direct control from a coding agent (research preview)

Let Claude Code, Codex or Cursor drive the computer and see the screen.

With the task tools, your coding agent asks Sai to do something and reads Sai's description of the result. With the control tools, your coding agent drives the computer itself: it runs Simulang blocks on it and gets each block's screenshot as an image its own model can see. See [Direct control](/documentation/concepts/direct-control) for the model.

Direct control is open to every account. The model calls the code makes (`callLLM`, `ground`, `stateSatisfies`) are billed to your account as usual.

## Setup

The control tools come with the Sai MCP server, next to the task tools; there is nothing to turn on. To leave them out, set `SAI_CONTROL=0` on the server.

In Codex, also set `tool_timeout_sec = 120` in the `[mcp_servers.sai]` table: Codex's default of 60 seconds is shorter than a block's default time limit.

## Tools

| Tool                    | Input                              | Output                                                                              |
| ----------------------- | ---------------------------------- | ----------------------------------------------------------------------------------- |
| `sai_control_open`      | `{ machine_id }`                   | `{ machine_id, expires_after_idle_s }`                                              |
| `sai_control_exec`      | `{ machine_id, code, timeout_s? }` | `{ result, logs, error?, variables, duration_ms }`, plus the screenshot as an image |
| `sai_control_interrupt` | `{ machine_id }`                   | `{ interrupted }`                                                                   |
| `sai_control_close`     | `{ machine_id }`                   | `{ ok }`                                                                            |

`machine_id` comes from `sai_machines`.

`timeout_s` defaults to 90 and is capped at 110, under the clients' tool timeouts. For a longer step, split it into several blocks.

The server also publishes the Simulang reference as the MCP resource `sai://simulang/reference`. The server's instructions tell the model to read it first. It is the same reference as [Simulang in direct control](/documentation/concepts/direct-control#simulang-in-direct-control).

## Try it

```
Take control of my Sai computer, open https://news.ycombinator.com in the browser, and tell me the top three stories. Close control when you are done.
```

The model opens control, runs a few blocks (open a tab, snapshot, read the outline), looks at the screenshots that come back, and closes control.

## Screenshots per client

| Client      | The model sees the screenshot                                 |
| ----------- | ------------------------------------------------------------- |
| Claude Code | Yes.                                                          |
| Codex       | Yes.                                                          |
| Cursor      | Yes with Claude and Gemini models. GPT models may not see it. |

Where the model does not see images, it can still read the screen as text: `page.snapshot()` and `app.snapshot()` return outlines with refs it can act on.

## Gotchas

- Control holds the computer. Your Sai tasks on it wait until the model closes control, or until it sits idle for 10 minutes.
- The control tools address the computer by `machine_id`. They have nothing to do with the `session_id` that `sai_task_start` returns.
- A free plan's computer is deleted when control closes.