Hosted computer driver

​

1. Get a key with the computers scope

export ALLTERNIT_API_KEY=alt_test_…

• AllternitComputers, a plain client for the /v1/computers routes.
• Anthropic SDK toolset subclasses for the computer and browser toolsets (BetaAbstractComputerToolset20260801, BetaAbstractBrowserToolset20260801).
• An OpenAI computer-use adapter and a Gemini computer_use adapter.
• The contract-v2 structured members (read_ui, act, run_batch, verify, request_human, use_credential, run_subtask, run_parallel, run_skill, skills) as typed client methods and as the computer_v2 function tool in every adapter.

​

2. Create a computer

curl https://api.allternit.com/v1/computers \
  -H "Authorization: Bearer $ALLTERNIT_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"name": "checkout-test"}'

​

3. Drive it from the Anthropic SDK

import Anthropic from "@anthropic-ai/sdk";
import { AllternitComputers, AllternitComputerToolset } from "@allternit/computer-driver";

const computers = new AllternitComputers(); // reads ALLTERNIT_API_KEY
const computer = await computers.create({ name: "checkout-test" });

const toolset = new AllternitComputerToolset({ client: computers, computerId: computer.id });
// Pass `toolset` in the tools list of the Anthropic SDK tool runner, then run your loop.

from allternit_computer_driver import AllternitComputers, AllternitComputerToolset

computers = AllternitComputers()  # reads ALLTERNIT_API_KEY
computer = computers.create(name="checkout-test")

toolset = AllternitComputerToolset(client=computers, computer_id=computer.id)
# Pass `toolset` in the tools list of the Anthropic SDK tool runner, then run your loop.

​

4. Or use the OpenAI or Gemini adapter

• OpenAI computer use: for each computer_call in the response, pass the call id and action to the adapter, then send its output back as computer_call_output. Set display to the display_width / display_height you declared on the tool.
• Gemini computer_use: for each function call, pass it to the adapter and send back the function response it returns.

import { AllternitComputers, runOpenAIAction } from "@allternit/computer-driver";

const computers = new AllternitComputers();
const step = await runOpenAIAction(
  { client: computers, computerId: "cmp_123", display: { width: 1280, height: 800 } },
  call.call_id,
  call.action,
);

​

5. Structured UI driving (contract v2)

import { AllternitComputers, ComputerV2Driver, computerV2AnthropicTool, runAnthropicComputerV2 } from "@allternit/computer-driver";
import { AllternitComputerToolset } from "@allternit/computer-driver/anthropic";

const computers = new AllternitComputers();
const v2 = new ComputerV2Driver({ client: computers, computerId: "cmp_123", onApproval: async (a) => askAPerson(a) });

const ui = await v2.read_ui({ app: "Safari" });
const field = ui.elements?.find((e) => e.role === "textfield");
if (field) await v2.act({ id: field.id, op: "set_value", value: "[email protected]", version: ui.version });

const sub = await v2.run_subtask({
  goal: "Fill the signup form and submit it",
  inputs: [{ name: "email", value: "[email protected]" }],
  success: [{ role: "button", name: "Submit" }],
});
// In an Anthropic loop, register computerV2AnthropicTool() next to the
// toolset in tools, and answer computer_v2 tool_use blocks with
// runAnthropicComputerV2({ client: computers, computerId: "cmp_123" }, name, input).

from allternit_computer_driver import AllternitComputers, ComputerV2Driver
from allternit_computer_driver.anthropic import AllternitComputerToolset, ComputerV2Tool

computers = AllternitComputers()
v2 = ComputerV2Driver(computers, "cmp_123", on_approval=lambda a: ask_a_person(a))

ui = v2.read_ui({"app": "Safari"})
field = next((e for e in ui.get("elements", []) if e.get("role") == "textfield"), None)
if field:
    v2.act({"id": field["id"], "op": "set_value", "value": "[email protected]", "version": ui["version"]})

sub = v2.run_subtask({
    "goal": "Fill the signup form and submit it",
    "inputs": [{"name": "email", "value": "[email protected]"}],
    "success": [{"role": "button", "name": "Submit"}],
})
# In an Anthropic loop, register ComputerV2Tool(computers, "cmp_123").to_dict()
# next to the toolset in tools.

​

Subtask safety statuses

• done — the goal is met.
• escalated — handed back with a reason and the current screen; continue from there yourself.
• needs_confirmation — the next step (held_step) needs the person’s confirmation; run it yourself with act/run_batch and the server asks them.
• paused — the safety monitor paused the subtask; stop and call request_human.
• denied — the step isn’t allowed on this computer (app/domain lists or a credential binding); find another way.
• use_api — make the named API/MCP call yourself; the screen is untouched.
• failed — ran out of steps, budget, or a hard error. Skipped parallel subtasks (budget exhausted) come back as skipped.

​

Structured-member errors

​

6. Or connect over MCP

{
  "mcpServers": {
    "allternit-computers": {
      "command": "computers-mcp",
      "env": { "ALLTERNIT_API_KEY": "alt_test_…" }
    }
  }
}

​

Approvals

{
  "error": { "type": "conflict_error", "code": "approval_required", "message": "…", "param": null },
  "approval": {
    "id": "apr_…",
    "member": "left_click",
    "toolset": "computer",
    "risk": "medium",
    "approve_url": "/v1/computers/cmp_123/approvals/apr_…"
  }
}

1. Get it approved. The project owner can approve in the console. If the project’s approval_mode is api_key, your key can call POST /v1/computers/{id}/approvals/{approval_id}; otherwise that call answers 403 approval_requires_owner.
2. Resend the same toolset call with approval_grant set to approval.id.

​

Limits

• Project flag. hosted_driver_enabled is set by Allternit. Off means 404 hosted_driver_disabled on every /v1/computers route.
• Per-key concurrency. Each key can have a set number of running computers (per_key_concurrency in computer settings). Past it, creating or starting a computer answers 429 concurrency_limit.
• Computer size. Each computer is a Linux computer with 1 vCPU, 2 GB of memory and a 10 GB disk. It has no desktop session; the screen (1280×800) starts with the first computer toolset call. The browser toolset runs headless Chrome.
• Pool capacity. A fixed number of developer computers can be running at once across all projects. When they’re all in use, creating or starting a computer answers 503 computer_capacity. Nothing is created and nothing is billed; retry in a few minutes, or stop a computer you aren’t using.
• Spend cap. When the project reaches its spend cap, calls answer 402 spend_cap_reached. See Billing and spend caps.
• Control. A computer that someone else is controlling answers 423 computer_controlled_elsewhere.
• Network egress. Computers created by outside developers reach the public internet only. Private address ranges, the cloud metadata address 169.254.169.254 and the host they run on are blocked.

​

Usage and pricing