Evaluation record · mcp-server-browserbase

Browserbase MCP Server

vhosted (mcp.browserbase.com) / @browserbasehq/mcp

Browserbase

MCPbrowser-automationcloud-browsersstagehandweb-scraping

Strong

About This MCP

Official Browserbase MCP server for cloud browser automation, powered by Stagehand v3: navigate, act, observe, extract (including iframes and shadow DOM), screenshots, and multi-session management. Hosted at https://mcp.browserbase.com/mcp with Browserbase covering Gemini costs, or local via @browserbasehq/mcp (API key + project ID). Stagehand v3 adds 20-40% faster automation via caching. Agent-driven browsing of arbitrary sites carries prompt-injection and credential-handling risk.

Last Evaluated: July 9, 2026

Official Website

Trust Vector Analysis

Dimension Breakdown

🚀Performance & Reliability

automation accuracy

Review of Stagehand's natural-language action model and v3 accuracy improvements

Evidence

Browserbase MCP changelog — Stagehand translates natural-language act/observe/extract instructions into resilient browser actions, with improved schemas in v3 for more intuitive data extraction

highVerified: 2026-07-09

operation success rate

Tool design review and hands-on interaction testing on common web flows

Evidence

browserbase/mcp-server-browserbase repository — Core tools (navigate, act, observe, extract, session start/end) build on Stagehand's self-correcting action layer; observe lets the agent find actionable elements before acting, improving success rates on dynamic pages

mediumVerified: 2026-07-09

complex page handling

Capability review of iframe/shadow-DOM extraction in Stagehand v3

Evidence

Stagehand v3 / Browserbase MCP setup docs — Stagehand v3 brings enhanced extraction across iframes and shadow roots plus 20-40% faster performance through automatic caching — handling page structures that break selector-based automation

highVerified: 2026-07-09

session stability

Session lifecycle and multi-session capability review

Evidence

Browserbase platform documentation — Cloud-managed browser sessions with keep-alive, proxies, stealth options, and configurable viewports; sessions survive client restarts and support parallel multi-session workflows

mediumVerified: 2026-07-09

error recovery

Error-path testing of failed actions and re-observation flows

Evidence

browserbase/mcp-server-browserbase repository — Failed actions return structured errors and the agent can re-observe the page; Stagehand's LLM-driven action resolution retries with fresh page state, though outcomes depend on the underlying model

mediumVerified: 2026-07-09

🛡️Security

prompt injection resistance

Threat modeling of untrusted web content entering both the client LLM and the Stagehand automation LLM

Evidence

Agentic browsing threat analysis — Extracted page content and observation results from arbitrary websites flow into the LLM context; in hosted mode page content is additionally processed by the Gemini model driving Stagehand. Malicious pages can embed instructions that hijack agent actions; no built-in content sanitization

highVerified: 2026-07-09

unauthorized action risk

Authorization boundary analysis of natural-language write actions

Evidence

MCP security guidance — act can perform any web action a browser allows (purchases, posts, form submissions); natural-language actions make precise pre-approval harder than explicit per-element tools, so host-level tool approval is the primary guardrail

mediumVerified: 2026-07-09

credential exposure risk

Analysis of credential and session handling in cloud-hosted browsers with recording enabled

Evidence

Browserbase platform documentation — Credentials typed during agent flows pass through cloud browsers and appear in session recordings/replays retained by Browserbase; injected cookies and persisted contexts extend logged-in session exposure beyond the local machine

mediumVerified: 2026-07-09

sandboxing isolation

Isolation review of per-session cloud browser architecture

Evidence

Browserbase platform documentation — Each session runs in an isolated cloud browser instance separated from the user's machine and profile — a materially better isolation story than driving a local browser with personal logins; SOC-2 Type 1 certified platform

highVerified: 2026-07-09

api key security

Credential mechanism review across hosted and local modes

Evidence

browserbase/mcp-server-browserbase README — Hosted endpoint authenticates via Authorization: Bearer BROWSERBASE_API_KEY header; local mode uses BROWSERBASE_API_KEY and BROWSERBASE_PROJECT_ID env vars. Keys are project-scoped but long-lived; no OAuth flow

highVerified: 2026-07-09

🔒Privacy & Compliance

browsing data exposure

Data flow analysis of extract/observe/screenshot outputs

Evidence

MCP data flow architecture — Extracted content, observations, and screenshots of every visited page (including authenticated content) are sent to the client's LLM provider as tool results

highVerified: 2026-07-09

session recording privacy

Privacy assessment of vendor-side session recording and retention

Evidence

Browserbase platform documentation — Browserbase records sessions for replay/debugging by default; everything the agent sees or types in the cloud browser — including sensitive form data — is captured in vendor-side session artifacts

highVerified: 2026-07-09

third party llm processing

Data sharing pathway analysis including the embedded automation LLM

Evidence

Browserbase MCP page — Hosted server runs Stagehand on Gemini (default google/gemini-2.5-flash-lite) with Browserbase covering the LLM costs — page content is processed by Google's model in addition to the user's own LLM provider; local mode allows custom model selection

highVerified: 2026-07-09

data residency control

Data residency review of the cloud-only browser execution model

Evidence

browserbase/mcp-server-browserbase repository — All browsing executes in Browserbase's cloud even with the local @browserbasehq/mcp package (which only relocates the MCP layer); no self-hosted browser-infrastructure option, so browsing data always transits vendor infrastructure

mediumVerified: 2026-07-09

👁️Trust & Transparency

documentation quality

Documentation completeness review

Evidence

Stagehand MCP setup docs — Clear documentation across browserbase.com/mcp, the GitHub README, and Stagehand docs covering hosted setup, local flags (proxies, stealth, keep-alive, viewport, model selection), and both transports

highVerified: 2026-07-09

open source transparency

Source code and license review

Evidence

browserbase/mcp-server-browserbase repository — Server open source under Apache-2.0 (~3.4k stars, 359 forks); Stagehand itself is also open source, though the browser infrastructure is a proprietary hosted service

highVerified: 2026-07-09

operation visibility

Logging and traceability assessment (the same recordings that raise privacy concerns aid auditability)

Evidence

Browserbase platform documentation — Session replays, live session view, and logs in the Browserbase dashboard provide unusually strong post-hoc visibility into exactly what the agent did in the browser

highVerified: 2026-07-09

vendor credibility

Maintainer reputation and project health analysis

Evidence

Browserbase — Well-funded, widely adopted browser-infrastructure startup; maintainer of Stagehand and an official MCP partner-ecosystem participant, though a younger vendor than Microsoft or Google

highVerified: 2026-07-09

⚙️Operational Excellence

ease of setup

Setup complexity assessment across hosted and local modes

Evidence

Browserbase MCP page — Hosted setup is a single URL (https://mcp.browserbase.com/mcp) plus an Authorization Bearer API key — Browserbase covers Gemini LLM costs on the hosted server; local mode is npx @browserbasehq/mcp with API key and project ID env vars

highVerified: 2026-07-09

feature coverage

Feature completeness assessment against browser-automation needs

Evidence

browserbase/mcp-server-browserbase README — Deliberately compact tool set (session start/end, navigate, act, observe, extract, screenshots, multi-session) — natural-language actions cover most flows but offer fewer explicit primitives than Playwright MCP's 25+ tools (no raw JS evaluation or network inspection)

highVerified: 2026-07-09

scalability

Scalability review of cloud browser infrastructure

Evidence

Browserbase platform — Cloud browser fleet supports parallel multi-session automation, proxies, and stealth at scales impossible with a single local browser

highVerified: 2026-07-09

maintenance activity

Repository activity and release-cadence analysis

Evidence

browserbase/mcp-server-browserbase repository — Actively maintained (~3.4k stars); tracks Stagehand releases, with the v3-powered hosted server and 20-40% performance gains shipped in 2026

highVerified: 2026-07-09

cost model

Cost analysis versus local browser automation alternatives

Evidence

Browserbase pricing — Free hosted-MCP Gemini inference lowers entry cost, but browser-session minutes are metered on Browserbase plans (free tier limited), unlike fully free local Playwright automation

mediumVerified: 2026-07-09

Strengths

+Stagehand v3 natural-language automation: act, observe, and schema-based extract including iframes and shadow DOM
+Hosted endpoint (mcp.browserbase.com) with Browserbase covering Gemini LLM costs — minimal setup
+Isolated per-session cloud browsers keep agent browsing off the user's machine and personal profile
+Scales to parallel multi-session automation with proxies, stealth, and keep-alive options
+Session replays and live view give strong auditability of agent actions
+Apache-2.0 open-source server (~3.4k stars) on the open-source Stagehand framework
+20-40% faster than prior versions via automatic caching

Limitations

!Page content flows to both the client LLM and the hosted Gemini automation model — a dual prompt-injection and data-exposure surface
!Session recordings capture everything the agent sees and types, including credentials, on vendor infrastructure
!Browsing always executes in Browserbase's cloud; the local package only relocates the MCP layer
!Natural-language act tool makes destructive web actions harder to precisely pre-approve
!Long-lived API keys with no OAuth flow
!Fewer explicit primitives than Playwright MCP (no raw JS evaluation or network inspection)
!Metered browser-session minutes beyond the free tier

Metadata

license: Apache-2.0

supported platforms

0: Any MCP client (hosted endpoint)

1: All platforms with Node.js (local @browserbasehq/mcp)

programming languages

0: TypeScript

mcp version: 1.0

github repo: https://github.com/browserbase/mcp-server-browserbase

github stars: 3400

package: @browserbasehq/mcp

remote endpoint: https://mcp.browserbase.com/mcp

api dependency: Browserbase cloud browsers + Stagehand v3

authentication: Browserbase API key (Bearer header on hosted; BROWSERBASE_API_KEY + BROWSERBASE_PROJECT_ID env vars locally); optional model API key for custom LLMs

automation llm: google/gemini-2.5-flash-lite by default (costs covered by Browserbase on hosted server); custom models supported locally

configuration flags

0: proxies

1: advancedStealth

2: keepAlive

3: viewport dimensions

4: context/cookie persistence

5: model selection

first release: 2025

maintained by: Browserbase

status: Active - Stagehand v3-powered hosted server

transport types

0: streamable-http (hosted)

1: stdio (local)

installation methods

0: remote-url

1: npx

2: npm

Use Case Ratings

code generation

Good for E2E testing and UI verification in clean cloud browsers, without local browser setup

research assistant

Strong for interactive web research at scale; exposed to prompt injection from untrusted pages

data analysis

Excellent structured extraction (schemas, iframes, shadow DOM) for scraping workflows

customer support

Can reproduce user-reported web issues with session replays as evidence

content creation

Useful for previewing and screenshotting published content

legal compliance

Vendor-side session recordings and multi-LLM data flow complicate confidentiality requirements

financial analysis

Capable scraper for public data; avoid authenticated financial workflows

education

Session replays make agent browsing behavior easy to demonstrate and teach

Similar MCPs

Playwright MCP

Microsoft

MCP Puppeteer Server

Anthropic (Archived)

Chrome DevTools MCP

Google (Chrome DevTools team)

Apify MCP Server

Apify