Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.getnen.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
Supported Models

ModelAPI String
Northstar (CUA, fast)tzafon.northstar-cua-fast
Northstar is Lightcone’s vision-language model purpose-built for computer use. It interprets screenshots and emits typed action items (click, type, scroll, etc.) that you execute against the desktop.

​
Integration Pattern

Northstar uses Lightcone’s Responses API with the built-in computer_use tool, not the JSON-Schema function-calling pattern that Anthropic and OpenAI chat models use. As a result, you do not consume Nen’s GET /tools JSON Schema — instead, you declare the screen size and Northstar emits action items in a fixed shape that you translate into Nen computer actions. 
from tzafon import Lightcone

lc = Lightcone()  # reads TZAFON_API_KEY

response = lc.responses.create(
    model="tzafon.northstar-cua-fast",
    tools=[{
        "type": "computer_use",
        "display_width": 1024,
        "display_height": 768,
        "environment": "desktop",
    }],
    input=[{
        "role": "user",
        "content": [
            {"type": "input_text", "text": "Open Firefox and go to wikipedia.org"},
            {"type": "input_image", "image_url": f"data:image/png;base64,{screenshot_b64}", "detail": "auto"},
        ],
    }],
)
Subsequent turns send the resulting screenshot back as a computer_call_output, chained via previous_response_id: 
response = lc.responses.create(
    model="tzafon.northstar-cua-fast",
    tools=[...],
    previous_response_id=response.id,
    input=[{
        "type": "computer_call_output",
        "call_id": call.call_id,
        "output": {"type": "input_image", "image_url": f"data:image/png;base64,{new_screenshot_b64}", "detail": "auto"},
    }],
)

​
Action Mapping

Northstar emits action items in response.output[*].action. Translate each one into the equivalent Nen computer tool action:
Northstar actionNen actionNotes
click (button: "left")left_clickcoordinate: [x, y]
click (button: "right")right_clickcoordinate: [x, y]
click (button: "wheel")middle_clickcoordinate: [x, y]
double_clickdouble_clickcoordinate: [x, y]
movemouse_movecoordinate: [x, y]
dragleft_click_dragUse path[0] for start_coordinate and path[-1] for coordinate
typetypetext: string
keypresskeyJoin keys with + (e.g. "ctrl+a"); map Control -> ctrl, Enter -> Return
point_and_typeleft_click + typeClick at (x, y) then type text
scrollscrollPick direction from the sign of scroll_y; amount = max(1, abs(scroll_y) // 100). Note: only up/down is supported — horizontal scroll returns HTTP 400. Drop scroll_x or skip pure-horizontal scrolls.
wait(sleep)No remote call needed
screenshotscreenshotAlready taken at end of every turn

​
Termination

Unlike Anthropic and OpenAI chat models — which signal completion by returning a plain text/message block instead of another tool call — Northstar tends to keep emitting computer_call items even after the visible task goal is reached. Plan for one of:
  • An external success oracle (URL check, DOM assertion, etc.).
  • A heuristic on message content: stop when the model finally emits a text reply with no computer_call.
  • A repetition detector: stop when the last 2 actions are identical and the screen is unchanged.
In practice, set --max-steps generously and treat hitting it as a normal exit, not an error.

​
Getting Started

Use nen example create to scaffold a ready-to-run agent project: 
nen example create lightcone-computer-use
cd nen-lightcone-cua
uv sync && cp .env.example .env
# Fill in NEN_API_KEY and TZAFON_API_KEY in .env
uv run python agent.py --task "Open Firefox and navigate to wikipedia.org"
See the Quickstart for a full walkthrough, or run nen example list to see all templates.

​
Next Steps

Claude (Anthropic)

JSON-Schema function-calling pattern with Claude

GPT (OpenAI)

JSON-Schema function-calling pattern with GPT

Computer Tool Reference

Full reference for all computer actions