Responses API

The Responses API lets you build a computer-use loop: Northstar looks at a screenshot, decides the next action (click, type, scroll), and you execute it, repeating until the work is done. Unlike Tasks where the model operates autonomously, the Responses API gives you control at every step.

How it works

1. Send input. Provide a text instruction and optionally a screenshot
2. Get output. The model returns either a text message or a computer_call action
3. Execute the action. Perform the click, type, or scroll on the computer
4. Feed back the result. Send a screenshot of the new state as computer_call_output
5. Repeat until the model returns a message (it’s done) or you decide to stop

Input format

The input field accepts either a string or an array of message objects:

• String. Simplest form, for text-only instructions
• Array. For images (screenshots) or structured multi-turn input

Both formats work identically for text-only requests. Use the array format when you need to attach a screenshot with input_image.

Create a response

The simplest form is a string as input:

from tzafon import Lightcone

client = Lightcone()

response = client.responses.create(
    model="tzafon.northstar-cua-fast-1.6",
    input="Open the terminal and check disk usage",
    tools=[
        {
            "type": "computer_use",
            "display_width": 1280,
            "display_height": 720,
            "environment": "desktop",
        },
    ],
)

import Lightcone from "@tzafon/lightcone";

const client = new Lightcone();

const response = await client.responses.create({
  model: "tzafon.northstar-cua-fast-1.6",
  input: "Open the terminal and check disk usage",
  tools: [
    {
      type: "computer_use",
      display_width: 1280,
      display_height: 720,
      environment: "desktop",
    },
  ],
});

When you need to include a screenshot (common in computer-use loops), use the array format:

response = client.responses.create(
    model="tzafon.northstar-cua-fast-1.6",
    input=[
        {
            "role": "user",
            "content": [
                {"type": "input_text", "text": "Click the search button"},
                {"type": "input_image", "image_url": screenshot_url, "detail": "auto"},
            ],
        },
    ],
    tools=[
        {
            "type": "computer_use",
            "display_width": 1280,
            "display_height": 720,
            "environment": "desktop",
        },
    ],
)

const response = await client.responses.create({
  model: "tzafon.northstar-cua-fast-1.6",
  input: [
    {
      role: "user",
      content: [
        { type: "input_text", text: "Click the search button" },
        { type: "input_image", image_url: screenshotUrl, detail: "auto" },
      ],
    },
  ],
  tools: [
    {
      type: "computer_use",
      display_width: 1280,
      display_height: 720,
      environment: "desktop",
    },
  ],
});

display_width and display_height are optional. If omitted and you include a screenshot in the input, the server infers the viewport dimensions from the image. If no image is provided, it defaults to 1024x768.

Process the output

The response output is an array of items. Each item has a type:

Type	Meaning
computer_call	The model wants to perform an action (click, type, scroll, and more)
message	The model is responding with text (work may be done)
reasoning	Internal reasoning (if available)

When you get a computer_call, the action field tells you what to do:

for item in response.output:
    if item.type == "computer_call":
        action = item.action
        print(f"Action: {action.type}")  # e.g., "click", "type", "navigate"
        print(f"Coordinates: ({action.x}, {action.y})")
        print(f"Text: {action.text}")

    elif item.type == "message":
        for block in item.content:
            print(block.text)

for (const item of response.output ?? []) {
  if (item.type === "computer_call") {
    console.log(`Action: ${item.action?.type}`);
    console.log(`Coordinates: (${item.action?.x}, ${item.action?.y})`);
    console.log(`Text: ${item.action?.text}`);

  } else if (item.type === "message") {
    for (const block of item.content ?? []) {
      console.log(block.text);
    }
  }
}

Action types

The model returns actions with coordinates in the 0–999 model space. Denormalize to pixel coordinates before passing to computer.click() or other methods: x = int(action.x / 1000 * display_width).

Here is the raw JSON on the wire, useful if you call the API directly without an SDK:

{
  "output": [
    {
      "type": "computer_call",
      "call_id": "call_abc123",
      "action": {
        "type": "click",
        "x": 450,
        "y": 320,
        "button": "left"
      }
    }
  ]
}

Note that the action kind is always action.type (e.g., "click", "type", "scroll") and coordinates are separate x and y fields, not a tuple.

The model can request these actions:

Action	Fields	Description
click	x, y, button	Click at coordinates
double_click	x, y	Double-click
triple_click	x, y	Triple-click (select a line)
right_click	x, y	Right-click
type	text	Type text
key / keypress	keys	Press key combination
key_down / key_up	keys	Hold / release key
scroll	x, y, scroll_y	Scroll vertically
hscroll	x, y, scroll_x	Scroll horizontally
navigate	url	Go to a URL (browser mode)
drag	x, y, end_x, end_y	Drag between two points
wait	:	Wait for the screen to settle
terminate	status, result	Work is complete (status: "success" or "failure")
answer	result	Answer a question with findings
done	text	Work is complete (alias)

Key names are lowercase. The model returns key names like "enter", "ctrl", "shift", "esc", not capitalized. The Computers API is case-insensitive, so you can pass action.keys directly to hotkey() without conversion.

Right-clicks use the button field. The model returns right-clicks as {"type": "click", "button": "right"}, not as a separate "right_click" action type. Check action.button to distinguish click types.

Extracting information

When tools includes computer_use, the model strongly prefers returning actions over text. To extract data from a page (for scraping, research, or monitoring), use a two-phase pattern:

1. Explore (with tools): the agent navigates, scrolls, dismisses popups
2. Extract (without tools): the model reads the screen and responds with text

If you include tools in an extraction request, the model will often scroll or click instead of answering your question. Omitting tools forces a text response.

TOOL = {
    "type": "computer_use",
    "display_width": 1280,
    "display_height": 720,
    "environment": "browser",
}

with client.computer.create(kind="browser") as computer:
    computer.navigate("https://example.com/pricing")
    computer.wait(3)

    # --- Phase 1: Explore WITH tools (agent scrolls, dismisses popups, etc.) ---
    screenshot_url = computer.get_screenshot_url(computer.screenshot())
    response = client.responses.create(
        model="tzafon.northstar-cua-fast-1.6",
        tools=[TOOL],
        input=[{
            "role": "user",
            "content": [
                {"type": "input_text", "text": "Scroll down slowly. Dismiss any popups. Stop when you can see pricing details."},
                {"type": "input_image", "image_url": screenshot_url, "detail": "auto"},
            ],
        }],
    )

    for step in range(10):
        computer_call = next(
            (o for o in (response.output or []) if o.type == "computer_call"), None
        )
        if not computer_call:
            break
        action = computer_call.action
        if action.type in ("terminate", "done", "answer"):
            break

        # Execute the action (click, scroll, type, etc.)
        # ... see the CUA protocol guide for the full execute_action helper ...

        computer.wait(1)
        screenshot_url = computer.get_screenshot_url(computer.screenshot())
        response = client.responses.create(
            model="tzafon.northstar-cua-fast-1.6",
            previous_response_id=response.id,
            tools=[TOOL],
            input=[{
                "type": "computer_call_output",
                "call_id": computer_call.call_id,
                "output": {"type": "input_image", "image_url": screenshot_url, "detail": "auto"},
            }],
        )

    # --- Phase 2: Extract WITHOUT tools (forces a text response) ---
    screenshot_url = computer.get_screenshot_url(computer.screenshot())
    extraction = client.responses.create(
        model="tzafon.northstar-cua-fast-1.6",
        input=[{
            "role": "user",
            "content": [
                {"type": "input_text", "text": "What is the price shown on this page? Reply with just the dollar amount."},
                {"type": "input_image", "image_url": screenshot_url, "detail": "auto"},
            ],
        }],
        # No tools, so the model must respond with text, not actions
    )

    for item in extraction.output or []:
        if item.type == "message":
            for block in item.content or []:
                print(block.text)  # e.g., "$29.99"

const tool = {
  type: "computer_use" as const,
  display_width: 1280,
  display_height: 720,
  environment: "browser",
};

const computer = await client.computers.create({ kind: "browser" });
const id = computer.id!;

try {
  await client.computers.navigate(id, { url: "https://example.com/pricing" });
  await new Promise((r) => setTimeout(r, 3000));

  // --- Phase 1: Explore WITH tools ---
  let screenshot = await client.computers.screenshot(id);
  let screenshotUrl = screenshot.result?.screenshot_url as string;

  let response = await client.responses.create({
    model: "tzafon.northstar-cua-fast-1.6",
    tools: [tool],
    input: [{
      role: "user",
      content: [
        { type: "input_text", text: "Scroll down slowly. Dismiss any popups. Stop when you can see pricing details." },
        { type: "input_image", image_url: screenshotUrl, detail: "auto" },
      ],
    }],
  });

  for (let step = 0; step < 10; step++) {
    const computerCall = response.output?.find((item) => item.type === "computer_call");
    if (!computerCall) break;

    const action = computerCall.action!;
    if (["terminate", "done", "answer"].includes(action.type!)) break;

    // Execute the action (click, scroll, type, etc.)
    // ... see the CUA protocol guide for the full action switch ...

    await new Promise((r) => setTimeout(r, 1000));
    screenshot = await client.computers.screenshot(id);
    screenshotUrl = screenshot.result?.screenshot_url as string;

    response = await client.responses.create({
      model: "tzafon.northstar-cua-fast-1.6",
      previous_response_id: response.id!,
      tools: [tool],
      input: [{
        type: "computer_call_output",
        call_id: computerCall.call_id!,
        output: { type: "input_image", image_url: screenshotUrl, detail: "auto" },
      }],
    });
  }

  // --- Phase 2: Extract WITHOUT tools ---
  screenshot = await client.computers.screenshot(id);
  screenshotUrl = screenshot.result?.screenshot_url as string;

  const extraction = await client.responses.create({
    model: "tzafon.northstar-cua-fast-1.6",
    input: [{
      role: "user",
      content: [
        { type: "input_text", text: "What is the price shown on this page? Reply with just the dollar amount." },
        { type: "input_image", image_url: screenshotUrl, detail: "auto" },
      ],
    }],
    // No tools, so the model must respond with text, not actions
  });

  for (const item of extraction.output ?? []) {
    if (item.type === "message") {
      for (const block of item.content ?? []) {
        console.log(block.text); // e.g., "$29.99"
      }
    }
  }
} finally {
  await client.computers.delete(id);
}

For a complete working example that extracts structured data from multiple websites, see competitor_research.py in the examples directory.

Multi-turn chaining

Use previous_response_id to chain conversations without resending the full history:

# First turn
response = client.responses.create(
    model="tzafon.northstar-cua-fast-1.6",
    input="Open the file manager",
    tools=[{"type": "computer_use", "display_width": 1280, "display_height": 720, "environment": "desktop"}],
)

# Execute the action, take a screenshot, then continue
followup = client.responses.create(
    model="tzafon.northstar-cua-fast-1.6",
    previous_response_id=response.id,
    input=[
        {
            "type": "computer_call_output",
            "call_id": response.output[0].call_id,
            "output": {"type": "input_image", "image_url": screenshot_url, "detail": "auto"},
        },
    ],
    tools=[{"type": "computer_use", "display_width": 1280, "display_height": 720, "environment": "desktop"}],
)

// First turn
const response = await client.responses.create({
  model: "tzafon.northstar-cua-fast-1.6",
  input: "Open the file manager",
  tools: [{ type: "computer_use", display_width: 1280, display_height: 720, environment: "desktop" }],
});

const followup = await client.responses.create({
  model: "tzafon.northstar-cua-fast-1.6",
  previous_response_id: response.id!,
  input: [
    {
      type: "computer_call_output",
      call_id: response.output![0].call_id!,
      output: { type: "input_image", image_url: screenshotUrl, detail: "auto" },
    },
  ],
  tools: [{ type: "computer_use", display_width: 1280, display_height: 720, environment: "desktop" }],
});

System instructions

Use instructions to set a system prompt that guides the model’s behavior across turns:

response = client.responses.create(
    model="tzafon.northstar-cua-fast-1.6",
    instructions="You are operating a desktop computer. Be careful and verify each action before proceeding.",
    input="Find the system settings and check the display resolution",
    tools=[{"type": "computer_use", "display_width": 1280, "display_height": 720, "environment": "desktop"}],
)

const response = await client.responses.create({
  model: "tzafon.northstar-cua-fast-1.6",
  instructions: "You are operating a desktop computer. Be careful and verify each action before proceeding.",
  input: "Find the system settings and check the display resolution",
  tools: [{ type: "computer_use", display_width: 1280, display_height: 720, environment: "desktop" }],
});

The instructions persist across chained turns when using previous_response_id.

Set store: false to skip response persistence. Responses created with store: false cannot be referenced via previous_response_id.

Manage responses

Each response has a status field:

Status	Meaning
in_progress	The model is still generating output
completed	The response is ready
failed	Something went wrong
cancelled	You cancelled it via cancel()

# Retrieve a response
response = client.responses.retrieve("resp_abc123")
print(response.status)  # "completed", "in_progress", etc.

# Cancel an in-progress response
client.responses.cancel("resp_abc123")

const response = await client.responses.retrieve("resp_abc123");
console.log(response.status); // "completed", "in_progress", etc.

await client.responses.cancel("resp_abc123");

Models

Model	Purpose	Input	Output
tzafon.northstar-cua-fast-1.6	Our latest computer-use model. Recommended; pin this version in production.	$0.50/M tokens	$1.50/M tokens
tzafon.northstar-cua-fast	Unversioned alias of the previous generation. Prefer the pinned -1.6.	$0.50/M tokens	$1.50/M tokens
tzafon.sm-1	Text model specialized for low-level systems programming.	$0.20/M tokens	$0.30/M tokens

List the models available to your key at any time with client.models.list().

For a fully managed experience where you don’t need to control the loop, use Tasks instead. The Responses API is for when you need full control over every step.

See also

• Computer-use loop: full implementation guide
• Tasks: fully managed alternative where Northstar operates autonomously
• Computers: computer lifecycle and all available actions
• How Lightcone works: how the API layers relate