Polar Browser raises $5.7 million for hours-long work automation

Polar Browser raised a $5.7 million seed round led by Madrona to develop an AI browser that automates hours-long research, recruiting, sales and operations tasks inside users' existing web accounts. The product, built by San Francisco-based Recursive Intelligence, is a direct Chromium fork that has completed individual tasks lasting more than 15 hours and logged over 4.5 million browser actions in seven months. CEO Kevin Jiang, a former Perplexity AI search engineer and MIT computer science graduate, announced the financing Wednesday alongside the company's July 29th launch.

Polar Browser https://polarbrowser.com/ raised a $5.7 million seed round led by Madrona to develop an AI browser that completes hours-long research, recruiting, sales and operations tasks inside users' existing web accounts. Kevin Jiang @jiang kev https://x.com/jiang kev , who is building Polar, announced the financing Wednesday in a thread on X https://x.com/jiang kev/status/2082496145994248488?s=46 , alongside a July 29th company launch post. https://x.com/jiang kev/status/2082496145994248488?s=46 https://x.com/jiang kev/status/2082496145994248488?s=46 Jiang studied computer science at MIT and previously built AI search at Perplexity, according to his personal site https://kevinjiang.xyz/ . An MIT profile he wrote in 2022 says he had already raised $1.4 million for a crypto startup accepted into Y Combinator's S22 batch before returning to school. He was also a four-time member of the USA Taekwondo National Team and taught poker theory at MIT Sloan. Polar is a product of San Francisco-based Recursive Intelligence, whose launch post https://polarbrowser.com/blog/introducing-polar describes a founding team of three with experience at MIT, Y Combinator, Jane Street, Citadel, Apple, Perplexity and Modal. Recursive Intelligence named former GitHub CEO Thomas Dohmke @ashtom https://x.com/ashtom , Modal founder Erik Bernhardsson @bernhardsson https://x.com/bernhardsson , Fundamental's Robert Yang @GuangyuRobert https://x.com/GuangyuRobert and Nico Christie @nicochristie https://x.com/nicochristie among the round's angel backers. Madrona https://x.com/MadronaVentures led the round. The financing gives Jiang's team capital to compete in a browser-agent category already occupied by products from OpenAI, Anthropic, Perplexity, Manus and Browser Use. Polar is taking a work-automation approach: users describe a task, then agents navigate websites, click, type, use authenticated sessions and coordinate subagents to complete it. Rebuilding the agent around Chromium Recursive Intelligence released Polar's predecessor, Composer, in alpha on January 26th. Composer used Electron and could run multi-hour tasks, but the team later said that architecture made it difficult to preserve authenticated sessions, operate browser extensions, interact with complex interfaces and build persistent memory. On May 18th, Recursive Intelligence replaced Composer with Polar, a direct Chromium fork. Owning the browser layer gives Polar more control over how agents interact with websites and lets them work across tabs and logged-in applications. The current product is available on macOS, with Windows users directed to a waitlist. Recursive Intelligence says users have assigned the product more than 4.5 million browser actions over seven months. The company also says Polar has completed individual tasks lasting more than 15 hours. One unnamed customer reportedly cancelled its Clay subscription and saves more than 25 hours per employee each week. Those usage and savings figures are company-reported, and Recursive Intelligence has not published revenue or customer totals. Polar's pricing https://polarbrowser.com/pricing shows how Jiang plans to turn that usage into a business. A free plan includes a one-time allocation of 3,000 credits and a limited daily refill. Pro costs $20 per month, Max costs $100 per month and team plans start at $50 per seat per month. Longer agent runs consume more credits. The mix places Polar between a prosumer browser and workplace automation software. Its sales pitch depends on whether the agent can reliably finish repetitive work inside the same applications employees already use, without requiring companies to integrate a separate tool for each workflow. Polar's benchmark claim needs context Polar says it ranks first on the Odysseys and BU Bench V1 browser-agent tests, ahead of systems using OpenAI and Anthropic models. Its evaluation repository https://github.com/recursive-hq/polar-evals reports 196 perfect tasks out of 200 on Odysseys, or 98%, and 95 completed tasks out of 100 on BU Bench V1. The repository includes judge outputs and says Polar used a port of BU Bench's official judge. Those are company-run evaluations. The official Odysseys benchmark https://odysseysbench.com/ says the strongest model in its own evaluation reached 44.5% perfect-task success. Polar's 98% result should therefore be treated as a self-published benchmark run rather than an independently maintained ranking. Jiang also acknowledged in the announcement thread that benchmarks do not fully capture the daily requirements of knowledge workers. Reliability outside a controlled evaluation remains the commercial test. Browser agents must handle changing page layouts, authentication, ambiguous instructions and interruptions while taking actions inside sensitive business accounts. Polar's privacy policy https://polarbrowser.com/privacy says browsing history, cookies and passwords generally stay on the user's device, while prompts, screenshots, page context, tool calls and other task data may be sent to Polar's systems and model providers to execute and review agent runs. The seed round funds a direct bet that browser agents can progress from staged demonstrations to recurring work. Recursive Intelligence says it is hiring engineers, designers and go-to-market staff in San Francisco. Jiang's challenge is to turn Polar's early action count into evidence that small teams will trust an autonomous browser with workflows they currently perform themselves or buy from specialized software vendors.