Polar: an AI browser built to do hours of work for you
Polar: an AI browser built to do hours of work for you
A Chromium-based Mac browser whose agent clicks, types and navigates sites while logged in as you, and can run tasks for hours or on a schedule. Launched on 29 July 2026 with a $5.7M seed round led by Madrona.
Rufus's note (2026-11-05): Yet another browser. My questions: (a) why use a browser when you can use your AI agent? (there must be a decent answer) (b) why this one vs the others? — Rufus Pollock
Links
- Website: https://polarbrowser.com/
- Launch post: Introducing Polar, 29 Jul 2026
- Funding announcement: Kevin Jiang, 29 Jul 2026
- Evals: https://github.com/recursive-hq/polar-evals
Overview
Polar is made by Recursive Intelligence, a team of three from MIT, YC, Jane Street, Citadel, Apple, Perplexity and Modal. Some of them helped build Perplexity's Comet browser. Its pitch: "A vast proportion of knowledge work runs through the browser", so a good enough browser agent can automate much of it.
Main features: ask the agent in any tab; save any task as a workflow that runs hourly, daily or on a trigger; 100+ connectors (Gmail, HubSpot and so on); and you can watch and take over at any time. It claims 4.5 million actions taken for users, and runs of 15+ hours.
| Plan | Price | Credits |
|---|---|---|
| Free | $0 | 1,000 to start, +50/day |
| Individual | $20/month | 10,000/month |
| Teams | $25/seat/month | 10,000/member/month |
Source: polarbrowser.com/pricing, Nov 2026. macOS only for now.
The answers to Rufus's questions
Why a browser rather than your AI agent? The browser is where your logins already are. An agent that lives in your main browser can use any site as you, including ones with no API or connector, and you can see what it does and take over. That is the real case. But general agents now cover much of it: Claude in Chrome and ChatGPT's agent mode also drive your logged-in browser. So the dedicated-browser advantage is shrinking to long unattended runs and scheduled workflows.
Why Polar over the others? Its claim, "beats Anthropic and OpenAI on every major web agent benchmark", rests on one self-reported chart: 98.0 on BU Bench V1, against 60.8 for GPT-5.4 and 44.5 for Claude Opus 4.6. Those are models from early 2026, not the current ones, and the post doesn't say how they were run. Treat it as marketing until someone independent tests it. The genuine differentiators are scheduled workflows and long runs.
Verdict: worth a look if you have a repetitive web workflow (recruiting, sales ops) that your current agent can't run on a schedule. Otherwise, the agent you already use is probably enough. See moc-ai-browsers.