Build with Northstar

Quickstart →How it works

Everything you need to build with Northstar, from your first API call to production. Give it an instruction and it operates a cloud browser or desktop to carry it out:

from tzafon import Lightcone

client = Lightcone()

for event in client.agent.tasks.start_stream(
    instruction="Go to news.ycombinator.com and give me the top 5 story titles",
    kind="browser",
):
    print(event)

import Lightcone from "@tzafon/lightcone";

const client = new Lightcone();

const stream = await client.agent.tasks.startStream({
  instruction: "Go to news.ycombinator.com and give me the top 5 story titles",
  kind: "browser",
});

for await (const event of stream) {
  console.log(event);
}

Northstar reads the screen and picks each action; Lightcone runs the browser or desktop it works on.

The Lightcone API and SDK are in beta. APIs and behavior can change. Pin your SDK version and update deliberately. Questions or issues? Email support@tzafon.ai.

Get started

QuickstartGet Northstar running a task in under 5 minutes.How Lightcone worksThe mental model, the API layers, models, and vision vs. DOM.AuthenticationGet your API key and configure the SDK.CLIRun tasks and drive computers from your terminal.API ReferenceFull reference for every endpoint.

Evaluating? See how Lightcone compares to RPA tools and to building on frontier computer-use APIs yourself. Want it fully on-device instead? See Run Northstar locally on a Mac.

Three levels of control

Same platform, three ways in. Pick how much of the loop you want to own.

ManagedTasksDescribe an outcome; Northstar runs the whole loop and streams progress. Watch it live and replay any run.ControlledResponses APIOwn the loop: the model picks each action, you execute it and feed back a screenshot. Inject logic between steps.InfrastructureComputersJust the environment: cloud browsers and desktops for your own model or automation, no Northstar involved.

For text generation and tool calling, use the OpenAI- and Anthropic-compatible Chat Completions endpoint.

Models

RecommendedNorthstar CUA Fast 1.6Our latest computer-use model. Pin it in production.
tzafon.northstar-cua-fast-1.6PreviousNorthstar CUA FastUnversioned alias of the prior generation.
tzafon.northstar-cua-fastTextsm-1Text model for low-level systems programming.
tzafon.sm-1

See the full model list and pricing.

What people build with Lightcone

Turn software without an API into an APIOperate ERPs, EHRs, Citrix desktop apps, and government portals on a cloud desktop, and expose the workflow as a REST endpoint.Replace RPA bots that keep breakingSelector bots break when the UI shifts. Northstar reads the screen instead, so you describe the task once and stop maintaining recordings.Run back-office operationsReconciliation, order entry, claims, work-order triage: multi-step jobs across systems that don’t talk, verified before they’re marked done.Collect web data at scaleParallel browser sessions with persistent logins and proxies; monitor prices or competitors across sites that block ordinary scrapers.Test like a real userGive a plain-language test plan. Northstar clicks through the flows like a customer and reports what broke, with a screenshot trail.Put computer use inside your productShip agent features without building the infrastructure. Drive the loop yourself, or hand work to managed Tasks and stream it into your UI.

Cookbook

Patterns from real deployments. Adapt and test them before shipping; they are starting points, not drop-in code.

Wrap a legacy app in a REST APITurn a GUI-only application into an HTTP service.Warm sessions & self-healingKeep an authenticated session alive and recover automatically.Verified structured extractionRead data off any screen, reliably.

Going to production

• Logins, credentials, and persistent sessions: work behind authentication safely
• Running in production: pinning, idempotency, retries, warm sessions
• Observability: live view, screencast recording, traces
• Continuous learning (preview): deploy a model that learns from corrections in your own traffic
• Security and data handling · Errors · Limits and behaviors

Integrations

Lightcone works with the tools you already use.

MCP ServerGive any agent Lightcone as tools.PlaywrightHybrid selector + model automation over CDP.LangChainUse Lightcone as a tool in LangChain agents.Vercel AI SDKLightcone as a tool in Vercel AI agents.All integrationsBrowser-Use, Mastra, Kernel, and more.

Install the SDK

pip install tzafon

npm install @tzafon/lightcone

Using a coding agent?

Paste this prompt into Claude Code, Cursor, or Codex to set up Lightcone end to end:

Set up Lightcone (the computer-use API by Tzafon), following these steps:

1. Install the CLI:  curl -fsSL https://lightcone.ai/install | sh
2. Authenticate:     run `lightcone login` (or set TZAFON_API_KEY from
                     https://lightcone.ai/dashboard → API Keys → Create)
3. Install the SDK:  Python `pip install tzafon` · TS `npm install @tzafon/lightcone`
4. Verify it works:  lightcone run "go to example.com and tell me the page heading" --kind browser

Reference docs while you work: https://docs.lightcone.ai/llms.txt

Prefer MCP? Wire Lightcone into your agent with one command, lightcone configure-agents, then see CLI and MCP Server.

Having trouble? See the Troubleshooting guide.