Browser Use Review 2026: The 115K-Star Agent That Quietly Dropped Playwright
Browser Use is the largest browser-automation project JAVIS covers by GitHub stars — bigger than Playwright's own MCP and CLI repos combined. It's also not what most people assume: an MIT-licensed Python agent that plans its own multi-step browser actions, that stopped running on Playwright over a year ago in favor of raw Chrome DevTools Protocol, and whose Cloud pricing has changed twice in 2026 alone. This review covers what it is right now — the free library, the paid Cloud, the real risks, and where it actually fits next to a driven tool like Playwright MCP.
| Library license | MIT — free, open-source |
| Still runs on Playwright? | No — raw CDP since Aug 2025 |
| Current agent generation | V4 (scheduling, Agency skill) |
| Cloud pricing model | Pure pay-as-you-go credits |
| Funding | $17M seed, Felicis-led (2025) |
The part everyone associates with Browser Use is the part it walked away from.
"Ask most developers what Browser Use runs on and they'll say Playwright — that was true, and it's genuinely not true anymore. In August 2025 the team ripped out their own Playwright dependency and rebuilt on raw Chrome DevTools Protocol, in their own words because 'the needs of AI browser agents are much narrower than the entire surface area that playwright provides.' A year later it's on V4, with agents that can be scheduled and left to work proactively."
What makes this worth taking seriously isn't the star count, even though 115,743 is a real number worth noting. It's that this is the one product in this comparison that actually plans — you give it a goal, not a script, and it decides the steps. That's a fundamentally different risk and reward profile than a tool something else drives.
An open-source agent core, plus an optional managed layer on top.
The MIT-licensed Python library (pip install browser-use or uv add browser-use) is the free, self-hosted core: point it at any supported model — OpenAI, Anthropic, Google, local models via Ollama, or Browser Use's own bu-2 model family — and it plans and executes browser tasks from a natural-language goal. Browser Use Cloud is a separate, optional paid layer: managed browsers, proxies, and either self-hosted-style billing or a fully managed agent that handles the LLM calls for you.
The asterisk worth flagging up front: the vendor's own benchmark claims an 82% pass rate on its own "Internal Bench Hard" test set — a real, disclosed number, but a self-defined, self-run one, not an independent audit.
The biggest browser-automation repo JAVIS covers, and a real business behind it.
browser-use/browser-use, verified on GitHub, Sep 22, 2026.
Verified on GitHub, Sep 22, 2026.
Felicis-led, announced Mar 23, 2025.
15 reviews — small sample, disclosed as such.
| When | What happened | Source |
|---|---|---|
| Oct 31, 2024 | Repository created; project starts as an open-source agent-browser interface | GitHub |
| Feb 2025 | Launch HN post — real community praise and security concerns from day one | news.ycombinator.com/item?id=43173378 |
| Mar 23, 2025 | $17M seed round announced, led by Felicis Ventures | browser-use.com/posts/seed-round, TechCrunch |
| Aug 20, 2025 | Drops its own Playwright dependency, rebuilds on raw CDP | browser-use.com/posts/playwright-to-cdp, official |
| Aug 17, 2026 | V4 agents gain scheduling + the new "Agency skill" for proactive, approval-gated work | browser-use.com/changelog/17-8-2026, official |
The praise and the biggest worry both showed up on day one.
One commenter called it "the best thing to ever happen to the RPA world," while another warned that launching Chrome in remote-debugging mode "presents a huge security risk" when a personal, logged-in browser is connected.
Reviewers describe it as making automation "even simpler" than Selenium-based approaches, with a "multi-modal approach" that's "really effective at navigating web page structures."
On why Playwright was dropped: "the needs of AI browser agents are much narrower than the entire surface area that playwright provides" — a rare case of a vendor explaining an architecture reversal in public.
Free if you self-host. Pay-as-you-go if you don't want to.
| Path | Cost |
|---|---|
| Open-source library, self-hosted, your own model key | $0 to Browser Use; your model provider's own rate |
| Browser Use Cloud — self-hosted-style browsers | $0.02/hour + $5/GB residential proxy (or $0.20/GB with your own proxy) |
| Browser Use Cloud — fully managed Agents | Model cost + 20% service fee (e.g. $1 in tokens → $1.20) |
| Minimum top-up / free credits | $5 minimum (never expires); $15 free for eligible Google/GitHub/Microsoft signups |
| Enterprise | Custom — HIPAA support, zero data retention available |
| Lifetime spend | Concurrent sessions unlocked |
|---|---|
| $0 | 10 |
| $200 | 50 |
| $1,000 | 250 |
| $5,000 | 500 |
| $25,000 | 1,000 |
Start with whether you want a planner or a driven tool.
Browser Use is built exactly for this — start with the free, self-hosted library.
Browser Use Cloud's pay-as-you-go tier covers that, starting at $5.
Look at Playwright CLI or Playwright MCP instead — both reviewed separately.
Read the Security section below before connecting anything via remote debugging.
The newest capability is agents that work while you're not watching.
As of the Aug 17, 2026 changelog, V4 agents can be scheduled with pause/resume controls, and a new "Agency skill" lets an agent "proactively research, prepare work, and return with follow-ups, asking for approval only before an external action like sending or publishing." That's a meaningful shift from single-shot task execution toward something closer to a standing assistant.
- Model-agnostic: OpenAI, Anthropic, Google, local models via Ollama, or Browser Use's own bu-2 family.
- Approval-gated external actions in the Agency skill is a real, sensible default, not just marketing language.
- MIT core means no license fee for the self-hosted path, ever.
- The vendor's own 82% "Internal Bench Hard" figure implies roughly 1 in 5 failures on its own hardest self-defined tasks.
- A planning agent's behavior is inherently less predictable step-to-step than a scripted tool.
Goal in, plan out, browser acts, you review.
Official Browser Use video demonstrating the agent's capabilities.
Published by the official "Browser Use" YouTube channel (youtube.com/@Browser-Use).
The most Playwright-adjacent product in this comparison isn't built on Playwright.
For over a year, Browser Use ran on Playwright, same as most browser-automation tools. In August 2025 the team published their own account of ripping that out: Playwright's node.js WebSocket adapter added "a 2nd network hop," which "incurs a meaningful amount of latency when we do thousands of CDP calls." Talking to Chrome DevTools Protocol directly removed that hop.
Bring any model — or let Browser Use be the MCP tool itself.
Browser Use is explicitly model-agnostic, and the README also lists it as an installable MCP service (com.browser-use/browser-use) — meaning it can act as the browser-driving MCP tool inside another agent's workflow, in addition to running as its own standalone planning agent.
| Model source | Notes |
|---|---|
| OpenAI (GPT models) | Bring your own API key |
| Anthropic (Claude) | Bring your own API key |
| Google (Gemini) | Bring your own API key |
| Ollama / local models | Fully self-hosted, no external API cost |
| bu-2 family (incl. bu-2-0-mini-preview) | Browser Use's own proprietary model, optimized for browser automation, added to the model picker in the Aug 2026 update |
Official Browser Use video on Browser Infrastructure.
Published by the official "Browser Use" YouTube channel (youtube.com/@Browser-Use).
The hourly rate is the smallest line on the real bill.
| Cost component | Approx. rate |
|---|---|
| Browser time (self-hosted-style) | $0.02/hour, billed per minute, 1-minute minimum |
| Residential proxy | $5/GB (default) or $0.20/GB with your own proxy |
| Managed agent LLM tokens | Whatever your chosen model costs, plus a 20% service fee |
| Self-hosted, open-source path | $0 to Browser Use; your model provider's own rate only |
The single biggest concern was raised at launch, and it's still the right one to ask about.
People with a goal, not a script.
| Situation | Why Browser Use fits | What to reach for instead |
|---|---|---|
| Task is describable as a goal, not a fixed sequence of steps | The agent plans its own path | — |
| Want to self-host for free with any model | MIT core, model-agnostic, no seat fee | — |
| Don't want to manage browser/proxy infrastructure | Browser Use Cloud pay-as-you-go | — |
| Need deterministic, guaranteed-repeatable execution | An LLM planner isn't the right tool for this | Playwright CLI / Playwright MCP |
| Task touches a real, logged-in personal browser session | Read the security section first | — |
Four real gaps, not manufactured ones.
A subscription-tier model was referenced as recently as an April 2026 changelog and has since been replaced by pure pay-as-you-go -- plan for the model to keep evolving.
Remote-debugging exposure isn't a bug that got fixed; it's inherent to how this class of browser-control tool works. Treat it accordingly.
The vendor's own 82% "Internal Bench Hard" number implies real, non-trivial failure rates on hard tasks -- and it's a self-defined benchmark, not independently audited.
Some early commenters predicted a large cloud vendor could offer a competing managed version under the same permissive license -- a real structural risk of the MIT choice, not resolved one way or the other as of this check.
If a fully autonomous planner isn't what you need, here's how to think about it.
| If you mostly need... | Compare Browser Use with... |
|---|---|
| A driven command surface for a shell-based coding agent | Playwright CLI — reviewed separately |
| A driven tool surface for an MCP-native chat/IDE client | Playwright MCP — reviewed separately |
| Hybrid hand-written flows with AI filling in the flexible parts | Stagehand (stagehand.dev) |
| Another autonomous browser-agent framework to compare against | Skyvern (skyvern.com) |
No affiliate relationship shapes this review.
JAVIS has not established a publisher affiliate program with Browser Use. Every CTA here points to the official product, official pricing, or the open-source repository.
Go to the official page.
Open Browser UseRead the alternative that matches your real question.
Choose the next articleBrowser Use makes the most sense next to the driven-tool alternatives it's often compared against.
Questions people are actually searching right now
Is Browser Use free?
The open-source Python library is free and MIT-licensed, with no seat fee. Browser Use Cloud (managed browsers/agents) is a separate, optional paid product billed on a pay-as-you-go credit basis.
Does Browser Use still use Playwright?
No. It dropped its own Playwright dependency in August 2025 and now runs on raw Chrome DevTools Protocol, per its own official engineering blog post.
How much does Browser Use Cloud cost?
Pure pay-as-you-go: $5 minimum credit top-up (never expires), $0.02/hour browser time, $5/GB residential proxy, and model cost + 20% for fully managed agents. There is no subscription tier as of this check.
Is Browser Use safe to connect to my personal browser?
A real, structural concern was raised at the product's own launch: connecting via Chrome's remote debugging protocol can expose a browser with real, logged-in credentials. Treat any driven session like handing over an unlocked tab.
Who funds Browser Use?
A $17M seed round led by Felicis Ventures, announced March 23, 2025, with participation from A Capital, Nexus Venture Partners, Y Combinator, Paul Graham, Liquid2, SV Angel and Pioneer Fund. No Series A or later round was found as of this check.
What is the "Agency skill"?
A V4 capability (added Aug 17, 2026) that lets an agent proactively research and prepare work on a schedule, pausing to ask for human approval only before an external, consequential action like sending or publishing.
Is Browser Use better than Playwright MCP?
Different category, not a strict "better": Browser Use is a fully autonomous planning agent; Playwright MCP is a tool surface an MCP client drives. Pick based on whether you want the LLM deciding the steps or a client deciding them.
Where this came from, and when it was checked.
GitHub stats (stars/forks/license/issues): GitHub data for github.com/browser-use/browser-use, Sep 22, 2026.
PyPI package version: pypi.org/project/browser-use, read directly, Sep 22, 2026.
Pricing (credits, concurrency tiers, Enterprise): browser-use.com/pricing, official, read directly, Sep 22, 2026.
Architecture change (Playwright to CDP): browser-use.com/posts/playwright-to-cdp, official engineering blog, dated 2025-08-20.
V4 / Agency skill: browser-use.com/changelog/17-8-2026, official.
Funding ($17M, Felicis-led): browser-use.com/posts/seed-round, official, corroborated by TechCrunch.
Real screenshots/media: browser-use.com/og/home-og.png, opengraph.githubassets.com repo card.
Official videos: youtube.com/watch?v=8oDCPK7POkQ and youtube.com/watch?v=Hp6LvqyAB-Y, authorship confirmed on YouTube (author "Browser Use", channel @Browser-Use).
User sentiment: Hacker News Launch HN thread (news.ycombinator.com/item?id=43173378), Product Hunt reviews.
