Introduction

Harness

Harness is an AI that sits next to you, sees what you see, and brings in the agents you already use. It lives in a small window that stays with you while you work. Ask it something and it answers from what is on your screen, or hands the job to Claude Code, Codex, Grok Bot or Bankr when they fit. Every answer comes back signed by the agent that did it.

It runs local-first: your screen indexes, full history, and memory database live on your own machine. In the anonymous and private settings, selected evidence plus a compact, expiring log of recent screen and shared-tab audio changes may pass through Harness and the inference route you chose.

Why it is different

  • It sees your screen. Ask about what is in front of you and get an answer grounded in the pixels that are actually there, not a guess from a description you had to type. It also keeps a private timeline of your screen, so you can ask what you were looking at an hour or a day ago.
  • It works in your browser. With the extension, point at any part of a page to ask about it in place, or hand Harness a job on the page and it does it, asking when a choice is yours.
  • It runs your agents. Harness Relay is a small app on the computer where your agents live. Any of them can hand work to another, or to several at once. Coding work happens in its own worktree, and merging is your call.
  • It reaches you on your phone. Pair your phone with your computer once. It pings you when an agent needs you, and you answer or merge from there. Nothing runs on the phone.
  • It remembers you. Harness builds a private, on-device model of your preferences, your work, and your corrections, and gets sharper about you the longer you use it.
  • It is private by construction. Your phone and the browsers you use Harness in reach your agents through your computer, sealed end to end: Harness carries those messages and cannot read them. Work for agents in the cloud, like Bankr or Grok Bot, is read to be done and labelled as such. The models that read your screen run in your browser. The memory of what it has learned lives on your device. The raw screen stream and full history stay local. The anonymous and private settings may send selected evidence and an ephemeral text and audio change log; private e2e keeps that log on your device.
  • The AI is yours to choose. Run a model locally or on your own OpenAI-compatible endpoint, bring your own Venice, Bankr, OpenAI or Dot key, or use managed inference and pay as you go. Every account starts with $5 of managed inference, free.

Running Harness

  • Hosted. Open tryharness.ai in Chrome, Brave, Arc or Edge and you are working in seconds, free, with no sign-up to try it.
  • Your browser. The extension is coming to the Chrome Web Store.
  • Your agents. Run npx tryharness on the computer where your agents live. It connects that computer, sets up the agents it finds and, on a Mac, adds Harness to your desktop. See Harness Relay.
  • Self-hosting. Running the full stack yourself is planned, not yet available. See Self-hosting to follow along.

Start here

Getting started