Watch Sai work

From @simular-ai/sai-mcp 0.2.0, every task started with sai_task_start comes with a live view:

• the computer's screen, streamed while the task runs. It is view-only: nothing you do in the view reaches the computer.
• Sai's steps as it works.
• permission requests, which you can answer with Approve, Approve for this task or Deny.
• the result, and the last frame of the screen once the task ends.

Where it shows

Your agent	What you get
Claude desktop and claude.ai	The view appears in the conversation under the task, as an MCP App. Open in a window pops it out so it stays visible while you chat.
Codex app	The view appears in the conversation, but Codex keeps it offline, so the screen can't stream there. Open the watch_url link the agent posts; Codex opens it in its built-in browser.
Claude Code, Codex CLI	No inline view. Open the watch_url link in your browser.
ChatGPT, Cursor, VS Code	These render MCP Apps, so the view should appear inline. Not yet tested.

You can also follow any task in the Playground.

Watching, agent control and human input

What you need	Interface
Watch Sai work	The task's watch_url (view-only) or the Playground.
Have your coding agent click, type and inspect screenshots	Direct control, using sai_control_open and sai_control_exec.
Complete a sign-in or other link-only input	The actual approvalUrl returned by approvals.
Let a person use the mouse and keyboard	A live_url from sai_live_link (see Hand the screen to a person), or the Playground on the user's own cloud computer.

The watch_url view never takes input. The Playground takes the mouse and keyboard on the signed-in user's own cloud computers; a Tenant's computers stay view-only there. The control API's owner takeover rule governs which credential holds a lease, not which viewer takes input.

If the MCP tools have not loaded yet, the control HTTP API is still available. Lack of inline streaming is a viewer limitation, not a lack of agent control. Closing a direct-control lease deletes a free-plan computer, so it is not a way to preserve a screen for a human handoff.

The watch_url link

sai_task_start returns a watch_url, and the agent is told to give it to you as a link. It opens the same view as a page served by the MCP server on your own computer, at http://127.0.0.1:<port>.

• Your API key stays in the MCP server; the page never sees it, and you don't sign in.
• The link carries a random token for this server process. Treat it like a temporary password: anyone on your computer with the link can follow the task and answer its approvals. It stops working when the MCP server exits.
• Set SAI_LOCAL_VIEW=0 to turn the page off. watch_url is then left out.

One viewer at a time

A computer's screen streams to one viewer at a time. If you open the task somewhere else (another app, the Playground), the first view stops and shows Watch here to take the screen back.

On a free account, the screen can drop for a moment between Sai's steps while the computer is handed back and forth. The view reconnects on its own.

For API users

The view uses POST /v1/agents/machines/:machineId/live, which you can call directly to build your own viewer.

Hand the screen to a person

The live view above only watches. When Sai needs you to use the computer, for a sign-in or a captcha, your agent calls the sai_live_link tool, { machine_id? } → { machine_id, live_url, expires_at }, and gives you live_url: a page anyone with the link can open in a browser without a Simular sign-in. You can use the mouse and keyboard there; while Sai is working, the page pauses its own input. When you are done, tell your agent, and it sends Sai a message to carry on. Anyone with the link can use the computer until expires_at, so don't share it.

Without the MCP server, the same link is liveUrl from POST /v1/agents/machines/:machineId/live. See When Sai needs a person.