Getting started

Harness runs in desktop Chromium browsers: Chrome, Brave, Arc or Edge. There is nothing to install to try it.

1. Ask about your screen

Go to tryharness.ai, share a tab, a window or your whole screen, and ask about what is on it. Trying it is free, with no sign-up.

Sharing is your browser’s normal screen-share prompt. Stop sharing at any time; when you are not sharing, Harness cannot see your screen.

Sign in with your email to keep your conversations, build memory and use everything else. You get a link by email, no password to remember. Every account starts with $5 of managed inference, free.

3. Let the on-device models load

On first run, Harness loads its screen-understanding models directly in your browser. This is a one-time load per device (they are cached afterward). Capture and screen understanding run locally. A selected current or recalled frame may still be sent to the model you choose when you ask a question.

4. Start chatting

Open the Harness window and ask about whatever is in front of you. Some things to try:

“What is this error telling me?”
“Summarize what is on this page.”
“What was I looking at earlier this morning?”
“Flag me if this deploy goes red.”

Go further

Your browser. The extension lets you point at any part of a page to ask about it, and hand Harness jobs to do on the page. It is coming to the Chrome Web Store.
Your agents. On the computer where Claude Code, Codex or your other agents live, run npx tryharness. See Harness Relay.
Your phone. Pair it once and it pings you when an agent needs you. See Your phone.

Choosing how the AI runs

Harness reaches AI in whichever way you prefer, set in Settings → Inference → Models & routing:

Use managed inference (harness), pay as you go. Every account starts with $5 of managed inference free, and you can top up any time.
Bring your own key (Venice, Bankr, OpenAI or Dot) and spend on your own account.
Connect your own endpoint, like a model running on your own machine, in Settings → Inference → Custom endpoints. Model execution is local and free when the endpoint is local; hosted turn orchestration still transits Harness.

The same page sets your privacy setting: anonymous, private or private e2e. See Models for what each means, and Plans for how managed credit and the paid Resident plan fit together.

Introduction Overview