| HybridPublic codebase/docs describe a DOM/accessibility serialization pipeline plus screenshots for the agent (DOMWatchdog / DOM serializer + screenshot path). | MITRecorded as MIT from the public sources on this row. | BothOpen-source library runs locally with Chrome/CDP; Browser Use Cloud offers hosted agents API and stealth browsers. | YesHosted MCP at https://api.browser-use.com/v3/mcp (task-level tools); local stdio browser-harness-mcp for step-by-step control of logged-in Chrome. | User browser and runtimeLocal harness can drive the user's Chrome session; Cloud uses managed browsers/profiles and API keys. Exact secret-vault behavior varies by path — see OSS sensitive-data docs. | Cloud stealth browsers claim 90.5% pass rate on Stealth Bench V2 (100 bot-protected sites); CAPTCHA solving advertised for Cloud (reCAPTCHA; hCaptcha newer/not guaranteed); off in EU region per marketing page. | OSS free (MIT). Cloud: pay-as-you-go — $0.02/hour browsers + $5/GB residential proxy; agents billed model tokens +20% service fee + browser time/traffic; $1 free credit for eligible new signups. Source: browser-use.com/pricing (accessed 2026-10-19). | | 2026-10-19 |
 StagehandUnknown Browserbase | HybridDocs/marketing emphasize hybrid accessibility-tree trimming; observe/act/extract use a11y+DOM context; optional CUA/agent modes can use vision models when configured. | MITRecorded as MIT from the public sources on this row. | BothlocalBrowser.launch / CDP attach for local; browserbase.launch for Browserbase cloud (stealth, proxies, Model Gateway, server-side cache). | PartialCLI-agent MCP adapter exists in-repo (stdio: run/snapshot/screenshot) but docs call integrations experimental and not published as a standalone package. Browserbase also documents hosted MCP with Stagehand-style tools (navigate/act/observe/extract) — treat as Browserbase-hosted path. | Agent runtimeCaller supplies model API keys (or Browserbase Model Gateway). Browser session auth is whatever cookies/storage the local or Browserbase context holds; no first-party 'credential vault' product called out in Stagehand quickstart. | When run on Browserbase, inherits Browserbase stealth/proxy/session features (see Browserbase docs). Stagehand itself is an SDK, not a stealth browser vendor. | OSS SDK free (MIT). Local runs: you pay your LLM provider. Cloud path uses Browserbase billing (Browserbase API key) — Stagehand docs do not restate Browserbase $/hour on the Stagehand pricing surface; Browserbase rates unknown in this seed (check Browserbase pricing separately). | | 2026-10-19 |
| VisionGitHub/README: relies on Vision LLMs / computer vision to interact with sites rather than only brittle DOM/XPath scripts; Playwright-compatible SDK adds AI page commands (act/extract/validate). | AGPL-3.0Recorded as AGPL-3.0 from the public sources on this row. | BothSelf-host with own LLM keys (pip/Docker/K8s); Skyvern Cloud app + API + hosted MCP. | YesHosted MCP https://api.skyvern.com/mcp/ (OAuth or x-api-key); local stdio via skyvern run mcp; large tool surface with optional scopes. | Cloud vaultEncrypted credential vault; model sees placeholders, real values injected at fill time. Integrates Bitwarden, 1Password, Azure Key Vault, custom HTTP vault. Self-host: credentials stay on your infra except what goes to the LLM provider you configure. | Cloud plans advertise CAPTCHA solvers (basic on Hobby, advanced on Pro), residential proxy / geo options on paid tiers; self-host docs note Cloud proxy/CAPTCHA are not included when self-hosting. | OSS self-host free aside from your LLM/infra (AGPL-3.0). Cloud credit plans from public pricing: Free 5,000 one-time credits; Hobby $29/mo 30,000 credits; Pro $149/mo 150,000 credits; Enterprise custom. Source: skyvern.com/pricing and docs Billing page (accessed 2026-10-19). Note: pricing FAQ wording on 'run' vs 'credit' has varied across pages — quote the Billing table as primary. | | 2026-10-19 |
| DOMMCP server uses Playwright accessibility snapshots (structured a11y tree), not pixel/vision models, by design. | Apache-2.0Recorded as Apache-2.0 from the public sources on this row. | Self-hostnpm package @playwright/mcp runs locally (npx); drives local browsers. No first-party hosted agent cloud as part of this package. | YesThis product is an MCP server. Docs also mention WebMCP page-registered tools and a separate Playwright CLI+SKILLS path for coding agents. | User browserUses the browser Playwright launches/attaches on the machine running the MCP server; logins are whatever is in that browser profile/context. | No dedicated stealth/CAPTCHA product claimed in Playwright MCP README; inherits ordinary Playwright browser fingerprint behavior unless you add your own tooling. | Free open source (Apache-2.0). You pay for the LLM/MCP client separately. No Playwright MCP cloud price page. | | 2026-10-19 |
 Claude in Chrome Anthropic | HybridHelp Center: extension reads page text, can take screenshots, uses debugger permission to click/type/screenshot, and can read console/network/DOM state for debugging. Exact internal a11y-vs-vision mix for every action is not fully specified in consumer docs — marked hybrid. | ProprietaryRecorded as proprietary from the public sources on this row. | CloudChrome extension product tied to Claude paid plans; drives the user's Chrome. Claude Code / Cowork can remote-drive the same extension. Cowork also rolling a built-in desktop browser option. | NoNot positioned as an MCP server. Integrates via Chrome extension + Claude Code/Cowork connectors. (Anthropic has other MCP connectors elsewhere; this browser product is extension-based.) | User browserShares the user's Chrome login state. Optional 1Password for Claude (beta, macOS): user approves biometrics; Claude never sees password/OTP per Help Center. CAPTCHA/login pages: Claude Code Chrome docs say it pauses for the user to handle. | No public stealth benchmark claimed. Safety classifiers gate actions; prohibited actions include purchases, creating accounts, handling sensitive card/ID data, etc. Site allow/block lists for Team/Enterprise. | Requires Claude paid plan (Pro, Max, Team, or Enterprise). No separate per-browser-hour price published on the Claude in Chrome help pages; usage is under Claude subscription (plan prices not restated here — see claude.com/pricing). | | 2026-10-19 |
 NotteUnknown Notte (nottelabs) | UnknownPublic marketing describes AI browser agents / Playwright-based automation; exact default perception channel (DOM vs vision vs hybrid) not clearly labeled on homepage/pricing as of access date — left unknown. | SSPL-1.0Recorded as SSPL-1.0 from the public sources on this row. | BothCloud sessions/agents/functions; docs/marketing also mention on-prem for Enterprise. Open-source SDK components on GitHub under SSPL. | YesMCP server documented (docs.notte.cc/integrations/mcp); Notte Browser MCP via pip install notte-mcp; Agent Skill and CLI also published. | Cloud vaultSecret Vaults: encrypted credentials; placeholder substitution outside model context on supported fill actions. Personas for digital identities (email/phone/credentials). | Cloud browser sessions marketed with anti-detection, residential proxies, and CAPTCHA solving. | Plans grant credits + usage rates from notte.cc/pricing (accessed 2026-10-19): Browser Sessions $0.01/hour; Browser Functions $0.05/hour; Residential proxies $10/GB (BYO proxy $0.20/GB on Startup); Free $10 one-time credits; Developer $20/mo; Startup $100/mo; Enterprise custom. LLM tokens: 'No Notte markup when using Notte-managed keys' per pricing page. | | 2026-10-19 |
 OpenAI Computer Use / Browser toolsUnknown OpenAI | HybridComputerToolset path is screenshot + mouse/keyboard (vision). BrowserToolset path exposes page text/controls as function tools (DOM-like structured reading). App chooses the path. | Proprietary APISeed value proprietary_api_plus_sample_app. The sample-app LICENSE file was not read in this pass. | BothYou supply the browser/desktop driver (local Playwright sample, Browserbase/Stagehand, Browser Use, Daytona, E2B, etc.). Separate Agents API path for OpenAI-hosted browser environment documented elsewhere. | PartialDocs say you may use remote MCP tools as your own UI tools; Computer Use itself is Responses API tools, not a dedicated Playwright-style MCP product. | UnknownApplication owns session, permissions, and cleanup. Guidance: keep credentials out of page text and screenshots; confirm consequential actions. Hosted partner browsers have their own auth models. | Not a stealth product; depends entirely on the environment/driver you attach. | API token usage (model + image input for screenshots) per OpenAI pricing; environment partners may charge separately. No single 'computer use plan' price restated on the guide page — see platform.openai.com pricing. Sample app is open for integration examples (license of sample app: check repo LICENSE). | | 2026-10-19 |
