Best AI Browser Automation Frameworks for Developers, Ranked by Task Success and Production Fit
We evaluated five browser-agent stacks on the same WebVoyager task set, scoring each on task success, form-filling reliability, production infrastructure, developer experience, and cost.
Browser Use is the highest-scoring open-source framework in the field, with a vendor-reported 89.1% success rate on WebVoyager and 91k+ GitHub stars as of mid-2026. Stagehand is the pick for TypeScript teams that want AI plus deterministic code in the same script; Skyvern wins when the workload is form-heavy or compliance-bound; Browserbase is the managed-runtime layer to pair with any of the frameworks above; Anthropic Computer Use is the fallback when the target is a canvas-only or vision-first surface the DOM stacks can't reach.
Five browser-automation stacks, one benchmark set, one ranking. We picked the frameworks and runtimes that engineering teams actually shortlist when they need an agent to log into a real site, fill a real form, or extract structured data from a page that changes without warning.
The category splits on two axes: DOM-driven vs vision-driven, and framework vs managed runtime. Browser Use, Stagehand, and Skyvern are the frameworks; Browserbase is the managed-browser runtime the others deploy against; Anthropic Computer Use is the vision-driven alternative for surfaces the DOM stacks can't reach. We report each on the same test plan below, with cost tracked alongside but kept out of the quality score.
We evaluated each framework at its default configuration on the WebVoyager task suite as reported by the maintainers, cross-referenced with the independent Notte multi-run reproduction and vendor documentation captured in July 2026. Task success and form-filling are vendor-reported or third-party benchmark figures; production infrastructure, developer experience, and cost are scored against documented plan tiers, SDK surface, and per-run economics.
Vendor-reported and third-party success rate on the WebVoyager benchmark, 586 diverse web tasks across live sites, evaluated with an independent LLM judge on agent actions and screenshots. Where a vendor's headline number could not be reproduced by an independent evaluator, we use the reproduction figure and note the gap. Weighted 30%.
Scored on the same WebVoyager subset plus category-level results on multi-step 'WRITE' tasks (logins, form fills, downloads) that are the primary use case for RPA-adjacent workloads. Skyvern published a 85.85% WebVoyager result and is the vendor-declared leader on WRITE tasks specifically. Weighted 20%.
Scored on the presence and quality of features that determine whether an agent survives a production workload: managed browser sessions, stealth/fingerprint handling, CAPTCHA solving, proxy support, session replay, concurrent-session limits, and stated compliance posture (SOC 2 Type II, HIPAA). Each capability was scored present-and-good, present-but-weak, or absent. Weighted 20%.
Scored on SDK ergonomics, language coverage, primitive design (act/extract/observe/agent vs single-goal agent), interop with Playwright, action caching, and openness of the license. We wrote the same three tasks (login-and-extract, multi-step form fill, structured data extraction) in each SDK and scored setup time, iteration speed, and debuggability. Weighted 20%.
Effective dollar cost per completed WebVoyager-scale task at each vendor's entry paid plan, calculated from published 2026 pricing pages and reproduction runs where available. For frameworks that require BYO LLM, we assume GPT-4o at the token counts reported in the Notte reproduction (7.4M input + 145K output tokens per 30-task WebVoyager30 × 8-run pass, ~$20 total). Reported alongside the quality score, never folded into it. Weighted 10%.
Browser Use is a Python-first open-source framework that parses the interactive elements of a rendered page (buttons, links, fields, dropdowns), feeds a simplified representation to an LLM, and executes the model's action plan against a headless browser. The maintainers report 89.1% on WebVoyager across 586 tasks, the highest published result among open-source browser agents. An independent multi-run reproduction by Notte on WebVoyager30 couldn't replicate the number and measured 77.3% self-reported / 60.2% LLM-verified, so treat the headline as a best-case ceiling rather than a production baseline. The project sits at 91k+ GitHub stars as of April 2026, is MIT-licensed, and works with any major model provider or a local Ollama endpoint. The trade-off is that the base library gives you a full-autonomy agent rather than the fine-grained code-plus-AI control that Stagehand offers.
Source: Browser Use (Muller/Zunic, YC W25) ↗Strengths
- Vendor-reported 89.1% on WebVoyager across 586 diverse web tasks
- 91k+ GitHub stars and MIT-licensed; largest community in the category
- Model-agnostic, works with OpenAI, Anthropic, Gemini, or local Ollama
Weaknesses
- Notte's independent reproduction on WebVoyager30 measured 60.2% LLM-verified, materially below the 89.1% headline
- Full-autonomy design offers less fine-grained control than Stagehand's primitives
How it scored, by metric
Stagehand is an open-source, MIT-licensed AI browser framework from the Browserbase team, distributed as a TypeScript SDK with a Python port and Stainless-generated clients for Go, Ruby, PHP, and Kotlin. Its four primitives (act, extract, observe, and agent) let developers choose per step whether to use AI or drop down to deterministic Playwright, and its v3 release replaced Playwright with a direct CDP engine that the maintainers report as 44% faster on shadow-DOM and iframe interactions. It runs locally against any Chromium browser and connects to Browserbase's cloud runtime with no code changes when a team is ready for managed sessions, Agent Identity, CAPTCHA solving, and session replay. The trade-off versus Browser Use is the smaller community (~22k GitHub stars, 700k+ weekly downloads) and the fact that the intended deployment path leads to Browserbase's paid tier.
Source: Browserbase ↗Strengths
- Four primitives give per-step choice between AI and deterministic code
- TypeScript-native with Python, Go, Ruby, PHP, and Kotlin SDKs
- Action caching replays known paths without further LLM calls
Weaknesses
- Community is smaller than Browser Use's (~22k vs 91k GitHub stars)
- Production path is optimized for Browserbase's paid cloud
How it scored, by metric
Skyvern combines computer vision, code execution, and LLM reasoning to automate browser workflows without hardcoded selectors, and it's explicitly optimized for form-heavy, portal-based work: logins, 2FA, CAPTCHA, downloads, procurement, invoice retrieval. The team reports 85.85% on WebVoyager 2.0 and is the declared leader on WRITE tasks, which is the category most enterprise RPA workloads actually live in. It's open source under AGPL-3.0 (~20k+ GitHub stars), self-hostable via Docker Compose, and the cloud product ships with SOC 2 Type II and HIPAA compliance for healthcare, insurance, and fintech workflows. The trade-offs are the AGPL license (public-service modifications must be open-sourced), a smaller general-web community than Browser Use, and credit-based pricing that gets expensive at high multi-step volume.
Source: Skyvern AI ↗Strengths
- Vendor-declared best-in-class on WebVoyager WRITE (form-filling) tasks
- SOC 2 Type II and HIPAA compliance for regulated workflows
- Route Memorization compiles successful AI paths into deterministic Playwright scripts
Weaknesses
- AGPL-3.0 license imposes source-sharing on public-service modifications
- Credit-based pricing at $0.05/step gets expensive at high multi-step volume
How it scored, by metric
Browserbase isn't a browser agent; it's the browser-agent infrastructure the other frameworks connect to for managed Chromium sessions, stealth fingerprinting, CAPTCHA solving, residential proxies, session replay, and observability. It raised a $40M Series B at a $300M valuation in June 2025 and now sells across a free tier, a $20/mo Developer plan (25 concurrent browsers, 100 browser hours), a $99/mo Startup plan (100 concurrent browsers, 500 browser hours, priority support), and a custom Scale plan. Overages run $0.12/hr (Developer) and $0.10/hr (Startup) on browser time, with additional per-call metering on Search ($7/1k), Fetch, and proxy bandwidth. It scores lower than the frameworks on pure task-success because it doesn't drive the agent loop itself; on infrastructure it's the clearest leader in the field.
Source: Browserbase, Inc. ↗Strengths
- Purpose-built stealth, proxy rotation, and CAPTCHA solving for adversarial sites
- Consumption pricing ($0.10–$0.12/browser hour on paid plans) scales predictably
- First-party Stagehand SDK plus works with Browser Use, LangGraph, and custom stacks
Weaknesses
- Not an agent, still requires a framework or custom code to drive the browser
- Multi-dimensional metering (hours, proxy GB, Search, Fetch, model tokens) is harder to forecast than flat-rate plans
How it scored, by metric
Anthropic's Computer Use is a vision-driven browser and desktop agent that operates on screen pixels rather than the DOM, reaching workloads the DOM-driven stacks can't (canvas-only apps, image UIs, aggressive anti-bot screens). Third-party benchmarking places it at ~78% reliability on common tasks, 12–17 percentage points behind the DOM-driven leaders on the same suite, but with a strictly wider addressable set of pages because it doesn't need selectors or a parseable element tree. It's the recommended fallback in composable production stacks: Stagehand or Browser Use for the 85–95% of pages the DOM stacks handle, Computer Use for the rest. The trade-offs are lower headline reliability, higher per-task cost (vision tokens are expensive), and Anthropic-provider lock-in.
Source: Anthropic ↗Strengths
- Only stack here that can drive canvas-only UIs and image-first pages
- Selector-free, resilient to DOM redesigns by construction
- First-party integration with Claude for reasoning-heavy agent loops
Weaknesses
- Third-party benchmarks put reliability 12–17 points behind DOM-driven stacks
- Vision-token cost is materially higher per task than DOM approaches
How it scored, by metric
The ranking above reflects five stacks evaluated against the same WebVoyager task set, at each vendor’s default configuration, with pricing verified against published July 2026 plan pages. The single largest separator at the top of the table isn’t raw task-success (the top three frameworks all sit inside a five-point band on their vendor-reported WebVoyager numbers) but how each stack handles the two problems that decide whether an agent survives production: page changes and adversarial anti-bot infrastructure.
What the scores measure
Task success carries the most weight because an agent that fails the benchmark fails the workload. We used WebVoyager because it’s the closest thing the category has to a shared reference, and because the Notte team published a reproduction methodology that lets us cross-check vendor-reported numbers against an independent multi-run evaluation. The gap between vendor and reproduction figures, most visible on Browser Use, where the maintainers report 89.1% and the reproduction measured 60.2% LLM-verified, is real and should shape how buyers read any single headline.
Where the field separates
Browser Use and Stagehand lead on general web tasks; Skyvern leads on form-filling. The difference is architectural: the WebVoyager benchmark tasks run on live websites in cooperative conditions, no aggressive bot protection, no Cloudflare, no DataDome, and the benchmark doesn’t measure speed, cost per task, or resilience to anti-bot measures. That single fact means a headline WebVoyager number can’t decide the pick for a workload that lives on protected production sites. On those workloads, the runtime (Browserbase’s managed stealth infrastructure, Skyvern’s built-in CAPTCHA and 2FA handling, or Anthropic Computer Use’s DOM-independent vision loop) matters more than the framework’s benchmark score.
Composition beats monoculture
The pattern that shows up in every production stack we surveyed is composition. Most production teams reach the same conclusion: you compose multiple tools, Stagehand for repeated workflows, Skyvern for the weird legacy bits, and something like a lightweight observation layer for the rest. The frameworks are building blocks, not religions. Browserbase sits underneath any of them as the managed runtime, and Anthropic Computer Use sits alongside as the vision-driven fallback for surfaces the DOM stacks can’t reach. A team standardizing on one tool for everything will pay for it in either coverage or maintenance; a team composing the right tool per workload will not.
Cost and license posture
Cost is tracked on the same runs but kept out of the quality score, because a buyer optimizing for spend and a buyer optimizing for reliability on Cloudflare-protected sites are answering different questions. The open-source frameworks (Browser Use MIT, Stagehand MIT, Skyvern AGPL-3.0) are free at the code layer and priced only by the LLM they call; Browserbase prices the managed runtime; Anthropic Computer Use prices the vision tokens. License posture matters for the AGPL-licensed Skyvern in particular, where public-service modifications must be open-sourced, a constraint that decides the choice for teams building customer-facing products on top of the framework.
- https://browser-use.com/
- https://stagehand.dev/
- https://www.skyvern.com/
- https://www.browserbase.com/
- https://www.anthropic.com/news/3-5-models-and-computer-use
- https://browser-use.com/posts/sota-technical-report
- https://github.com/nottelabs/open-operator-evals
- https://www.browserbase.com/pricing
- https://www.skyvern.com/pricing
Q.Which browser agent framework has the highest success rate on WebVoyager?
Browser Use reports the highest published open-source result on WebVoyager at 89.1% across 586 tasks, followed by Skyvern 2.0 at 85.85%. The caveat is reproducibility: an independent multi-run evaluation by Notte on the WebVoyager30 subset measured Browser Use at 77.3% self-reported and 60.2% LLM-verified, so the 89.1% figure is best treated as a best-case ceiling rather than a production baseline. Real-world success on Cloudflare- or DataDome-protected sites will be lower than any of these numbers.
Q.When should I pick Stagehand over Browser Use?
Stagehand is the right pick when you work in TypeScript, want per-step choice between AI and deterministic code, and need the same script to survive DOM redesigns without becoming a black-box agent. Its four primitives (act, extract, observe, agent) plus Zod schema validation give you the option to drop down to Playwright wherever precision matters, and its action-caching layer replays known paths without further LLM calls. Browser Use is the better pick for Python-first teams that want a full-autonomy goal-driven agent.
Q.Do I need Browserbase to run these frameworks in production?
No. Browser Use, Stagehand, and Skyvern all run locally against any Chromium browser at zero infrastructure cost beyond LLM tokens. Browserbase is what teams add when self-hosting Chromium fleets becomes operational toil: managed sessions, stealth fingerprinting, residential proxies, CAPTCHA solving, and session replay. Its Developer plan is $20/month with 100 included browser hours and $0.12/hr overage; the Startup plan is $99/month with 500 included hours at $0.10/hr overage.
Q.When is Anthropic Computer Use the right choice over a DOM-driven framework?
Computer Use is the right pick when the target surface is canvas-only, image-first, or aggressively anti-bot in a way that hides the DOM from parsers. Third-party benchmarking places it at ~78% reliability on common tasks, 12–17 percentage points behind DOM-driven stacks like Stagehand and Browser Use, and vision tokens make it more expensive per task. Most production teams treat it as a fallback layer in a composable stack: DOM-driven for the pages that parse cleanly, vision-driven for the rest.
Hana Koizumi evaluates image, audio, and agentic tool use. She writes the task suites that probe vision and function-calling reliability, and she scores how a product behaves when it has to act, not just answer.
Other leaderboards
- AI for Small and Mid-Size Business Best AI Marketing Automation Platforms for Small and Mid-Size Businesses, Ranked
- Productivity Best AI Content Optimization Platforms for SEO, Ranked by Recommendation Quality and Workflow
- Productivity Best AI Meeting Notetakers for Professionals, Ranked by Notes, Workflow, and Cost