# Bu · Browser Agent — element 11 of 58

> Hands for the open web. Turns websites without APIs into automated ones.

- **Group:** 3 · Agents & Automation
- **Necessity:** Optional
- **Price band:** $ · under $30/mo
- **Maturity:** Experimental
- **Edition:** v2026.Q3 · verified 2026-09-13

## Leading tools (v2026.Q3)

- **Claude in Chrome** — assisted browsing
- **Browser-use** — open-source control
- **Browserbase + Stagehand** — production browser infra
- **Perplexity Comet** — the free consumer agent
- **Playwright MCP** — the free dev default

## Our take

Powerful, brittle, improving monthly. Use for the long tail; use real APIs everywhere one exists.

## Combines with

Ag, Wf


## The top 5 — deep dossier (verified 2026-09-13)

Claude in Chrome for the human-in-the-loop case — it is the only assisted-browsing agent that ships published adversarial safety numbers (prompt-injection success cut 23.6%→11.2%, and 35.7%→0% on four browser-specific attacks) and org-level site allowlists, on every paid Claude plan. Browser-use when you are building the automation yourself — 107k GitHub stars, MIT, and the de-facto open-source standard. Browserbase + Stagehand when that automation has to run in production at volume (proxies, captchas, session replay). Comet when you want a free consumer agentic browser; Playwright MCP when the browser work is your own app — testing, verification, coding-agent loops — where it is free and deterministic. The category consolidated hard in mid-2026: OpenAI is retiring Atlas (Aug 9), Google folded Mariner into Gemini-in-Chrome, Microsoft folded Copilot Mode into Edge. The standalone AI browser is dying; the browser agent is not.

1. **Claude in Chrome** (Anthropic) — Included in paid Claude plans (Pro $17–20/mo · Max $100–200/mo · Team/Enterprise) · beta. Best for: Founders and teams who want an agent working inside their real, logged-in browser — inbox triage, form filling, dashboard extraction, scheduled recurring tasks — with admin-managed site controls. Why: The safety-engineering leader in a category defined by prompt-injection risk: Anthropic published red-team numbers (123 cases, 29 attack scenarios; success rate 23.6%→11.2% after mitigations, 35.7%→0% on four browser-specific attacks) and shipped org-wide allowlists/blocklists when it expanded from Max-only to all paid plans (Dec 18, 2025). It is also the best-integrated: hands for Claude Code (build in terminal, verify in browser) and Cowork, plus multi-tab work, background tasks, and cron-style scheduled browsing. Watch: Still labeled beta with 'unique risks' — Anthropic itself tells you to keep it away from financial transactions and passwords. Chrome-desktop-only, no API surface: this is an assistant, not infrastructure you can build a product on. [https://claude.com/claude-for-chrome](https://claude.com/claude-for-chrome)
2. **Browser-use** (Browser Use (YC W25, MIT)) — OSS free (MIT) + your LLM tokens · Cloud: browsers $0.02/hr metered by minute, hosted agents from $0.36/M input · $1.44/M output tokens. Best for: Engineers building their own browser automation — the open-source library that turns any LLM into a web agent, with a usage-priced cloud when you outgrow localhost. Why: The open-source standard: 107k GitHub stars (Aug 2026), MIT-licensed, Python + TypeScript, any model provider — it is the layer Manus and 20+ YC startups were built on. Its DOM-to-text approach (parse interactive elements instead of screenshots) made agents cheaper and more reliable than vision-first designs, and the repo claims #1 on the Odysseys long-horizon benchmark (87.4% across 200 web tasks). $17M seed led by Felicis (Mar 2025) with Paul Graham participating. Watch: Quality is whatever model you plug in, and long agent runs burn tokens fast. The cloud business is young and the space is crowded with better-capitalized infra (Browserbase); benchmark claims are self-reported. [https://browser-use.com](https://browser-use.com)
3. **Browserbase + Stagehand** (Browserbase) — Free tier · Developer $20/mo (100 browser-hrs, $0.12/hr after) · Startup $99/mo · Enterprise custom; Stagehand SDK free (MIT). Best for: Production browser automation at scale — headless browser fleets with proxies, captcha solving, session replay, and Stagehand's self-healing act/extract/observe primitives on top. Why: The infrastructure pick: $40M Series B at a $300M valuation (Jun 2025), ~50M sessions served in 2025, and marquee production users — Airtable built its Hyperagent platform on it (Jul 2026), Amplitude runs sales automation on it. Stagehand (MIT, ~23k stars, TypeScript + Python) is the most credible middle path between brittle Playwright selectors and fully autonomous agents, and the Jun 30, 2026 'Browserbase Agents' launch added fully managed prompt-to-agent runs. Watch: Costs compound at volume (browser-hours + proxies + agent runs are metered separately). The captcha/anti-bot arms race means reliability on hostile sites is never guaranteed, and its new managed Agents put it in competition with its own customers. [https://www.browserbase.com](https://www.browserbase.com)
4. **Perplexity Comet** (Perplexity) — Free · Comet Plus $5/mo · Pro $20/mo · Max $200/mo (background assistants) · Enterprise from $40/seat. Best for: A free, full agentic browser for research-heavy individual work — summarize, compare across tabs, handle email, shop — on Mac, Windows, iOS, and Android. Why: The distribution winner among standalone AI browsers — and after Atlas's retirement, effectively the last one standing at scale: free worldwide since Oct 2025, Android (Nov 2025) and iOS (Mar 2026) apps, enterprise MDM deployment since Mar 2026. Its assistant works inside your logged-in sessions, which is exactly why Amazon sued — making Comet the test case for whether user agents may act in authenticated accounts. Watch: The Amazon CFAA litigation (preliminary injunction Mar 2026, stayed pending Ninth Circuit appeal) could constrain logged-in commerce agents for everyone. Security researchers keep landing hits (CometJacking; a sub-4-minute phishing demo, Mar 2026). Consumer product, no developer API. [https://www.perplexity.ai/comet](https://www.perplexity.ai/comet)
5. **Playwright MCP** (Microsoft) — Free, Apache-2.0 · your LLM tokens only. Best for: Giving any MCP-speaking agent (Claude Code, Cursor, VS Code, Cline) deterministic browser hands — testing your own app, verifying agent-built UIs, automating cooperative sites. Why: The dev-loop default by sheer volume: 6.3M npm downloads per week (Aug 2026) and 35.5k stars. Its accessibility-snapshot approach — structured a11y trees instead of screenshots — is fast, cheap, and works without vision models, with 60+ tools covering navigation, network mocking, and trace recording. If your coding agent needs to see and click what it just built, this is the free, boring, correct answer. Watch: A tool server, not an agent — you bring the loop and the judgment. No stealth, proxies, or captcha handling, so it struggles on hostile third-party sites; snapshot-heavy sessions get token-expensive on long tasks (Microsoft itself notes CLI workflows are more token-efficient at high throughput). [https://github.com/microsoft/playwright-mcp](https://github.com/microsoft/playwright-mcp)

### How to choose
- If The site has an API or MCP server → Use it and skip this element — browser agents are for the long tail without APIs. They are slower, costlier, and more brittle than any real integration, and the vendor may sue (see Amazon v. Perplexity).
- If You want an agent doing your browsing — inbox, dashboards, forms, recurring chores — under your eyes → Claude in Chrome on your existing paid plan; keep 'ask before acting' on for anything sensitive and let admins pin the site allowlist.
- If You are building browser automation into your product → Prototype on Browser-use (free, MIT, any model); move to Browserbase + Stagehand when you need proxies, captcha solving, concurrency, and replay in production.
- If The browser work is your own app — E2E tests, verifying what your coding agent just built → Playwright MCP. Free, deterministic, no stealth needed on sites you control, and every coding agent already speaks MCP.
- If You automate other companies' logged-in sites at commercial scale → Get counsel before code. The Amazon v. Perplexity injunction fight (Ninth Circuit, argued Jun 11, 2026) is deciding whether authenticated agent access is a CFAA violation — and 4 of 7 agentic browsers failed a Jul 2026 same-origin security audit.

### The field (21 more)

Gemini in Chrome, ChatGPT Atlas (sunsetting), OpenAI Operator (dead), Project Mariner (renamed), Edge Copilot Mode (renamed), Skyvern, Steel, Hyperbrowser, Airtop, Lightpanda, Agent Browser, Amazon Nova Act, Manus Browser Operator (acquired), Dia (acquired), Opera Neon, Fellou, Genspark AI Browser, Firefox AI Window, Cloudflare Browser Run (renamed), MultiOn (renamed), Adept (acquired)

Full dossier data: https://elems.ai/e/bu.json

---
Source: [elems.ai](https://elems.ai/e/bu.html) — the periodic table of the AI-led startup. Data: https://elems.ai/elements.json (CC BY 4.0, cite elems.ai).