Sai is Simular's computer-use agent, available to developers as the Sai API. Give it a task in plain English and it does the work on a real computer the way a person would, by seeing the screen, clicking and typing: on a Windows or Linux cloud computer that Simular hosts, or on your own Mac or Windows PC. Your agent can then get things done where there is no API: web apps behind a login, admin consoles, desktop software, forms with no import. Run a workflow once and Sai can save it as code: later runs replay that code and ask the model only where the screen matters, which cuts tokens by up to 90% on repetitive tasks, and the agent takes over when the UI changes. On OSWorld 2.1, Sai has the highest partial score on the official leaderboard, 79.38%, and the only OSWorld 2.1 entry that publishes its cost, $14.34 per task, with about 1.5x fewer model calls than models used on their own, Oct 2026. Coding agents (Claude Code, Codex, Cursor) use it through the
@simular-ai/sai-mcpMCP server. The free plan includes a cloud computer.
POST /v1/agents/message over HTTPS, a Python SDK, and the @simular-ai/sai-mcp MCP server for Claude Code, Codex, Cursor and any stdio MCP client.Use Sai when:
Write code yourself when:
Most computer-use agents re-plan every step on every run. Sai can turn a workflow it has done into a reusable skill: the steps that stay the same become code, and the model is asked only about what depends on the screen.
Save this as a replayable skill named invoice_export, with vendor and month as parameters.Run invoice_export for vendor Acme, month March. Sai replays the code, and when a step breaks because the UI changed, the agent takes over from there.Open our staging site in Chrome, sign up with the test Gmail account, click the verification email, and tell me where the flow breaks.Log in to the vendor portal (it has no API) and list last month's invoices with their totals.Enter each row of the uploaded CSV into the supplier form at https://example.com/suppliers/new, which has no import, and report the rows that failed validation.Follow the steps in this bug report in Chrome on Windows and describe exactly what happens at each step.In the admin console, export the user list and tell me which accounts haven't signed in for 90 days.sapi_ and are shown once. Free accounts verify a phone number before their first key.sapi_YOUR_KEY with the key. Node.js 22.12 or newer is required.Use Sai to open Notepad on my cloud computer, type "Hello from Sai", and tell me the window title.claude mcp add sai -e SAI_API_KEY=sapi_YOUR_KEY -- npx -y @simular-ai/sai-mcp
Optional: install the Sai skill, which teaches Claude Code the task loop.
npx -y @simular-ai/sai-mcp init-claude
codex mcp add sai --env SAI_API_KEY=sapi_YOUR_KEY -- npx -y @simular-ai/sai-mcp
Add this to ~/.cursor/mcp.json (or .cursor/mcp.json in a project), then restart Cursor. The same block works for any stdio MCP client.
{
"mcpServers": {
"sai": {
"command": "npx",
"args": [
"-y",
"@simular-ai/sai-mcp"
],
"env": {
"SAI_API_KEY": "sapi_YOUR_KEY"
}
}
}
}
SAI_API_KEY as shown above.sai_machines, sai_models, sai_task_start, sai_task_wait, sai_task_approve, sai_task_abort, sai_upload.sai_control_open, sai_control_exec, sai_control_interrupt, sai_control_close. These let the coding agent operate a cloud computer and inspect screenshots. Read direct control and the control HTTP API before use. Missing MCP tools after setup do not prevent using HTTP directly.watch_url. Codex's inline view cannot stream; use that link instead. Read live viewing and the live-screen HTTP API. The MCP viewer and watch_url are view-only. For a person to use the mouse and keyboard, call sai_live_link and return its live_url. For sign-ins and other link-only human input, return the actual approval link.machine_id, not a task's session_id. Closing its lease deletes a free-plan computer; it is not a handoff that preserves the screen for the user.sai_task_start returns a session_id. Call sai_task_wait with it, passing back the cursor, while status is running. idle means text holds Sai's answer.needs_approval, decide with sai_task_approve. When the approval is link-only (sign-ins, credentials, phone checks), give approval_url to the user and keep waiting.The raw request behind every task. It streams events until finish. If Sai asks before a step, the stream sends data-approval-request and waits: when isLinkOnly is false, answer it from another terminal with POST /v1/agents/approve ({ "approvalId": "…", "response": "yes" }, or "task" to allow the rest of the task). Only link-only approvals are completed at approvalUrl.
curl -N https://api.simular.ai/v1/agents/message \
-H "Authorization: Bearer $SAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"message": "Open Notepad, type Hello from the Sai API, and tell me the window title"}'
The Sai API lets a program or a coding agent hand a task to Sai, Simular's computer-use agent, in plain English. Sai does the task on a real computer by seeing the screen, clicking and typing, and returns the result.
A browser automation API drives a web browser. Sai operates the whole computer, so one task can move between websites, desktop apps and file dialogs, on Windows, Linux or macOS.
A Windows or Linux cloud computer that Simular hosts, or your own Mac or Windows PC. On Pay as you go, up to 5 cloud computers can run at the same time.
Run a workflow once, then ask Sai to save it as a replayable skill. Later runs execute the saved code and ask the model only where the screen matters, which cuts tokens by up to 90% on repetitive tasks.
Sai pauses for approval before sensitive steps. Sign-ins, credentials and phone checks are link-only approvals: Sai sends the person an approval link, and they complete the step themselves.
On the official OSWorld 2.1 leaderboard, Sai has the highest partial score, 79.38%, at $14.34 per task, the only published cost on the leaderboard. On binary accuracy it is second, at 41.67%. Details: Sai on OSWorld 2.1.
Explore is free, with daily credits and limited access to a Windows cloud computer. Pay as you go is $50 a month, Sai Unlimited $500 a month, and Enterprise is custom. See pricing.
Yes, through the @simular-ai/sai-mcp MCP server. One command adds it; see Getting started above.
POST /v1/agents/message, the request behind every task