Browser Use
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
作者提供的说明(英文原文)
Browserbase MCP server: cloud browser sessions, agents, contexts, fetch, search, functions
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
BROWSERBASE_API_KEY, BROWSERBASE_REQUEST_TIMEOUT_MS, MCP_WORKSPACE_PATH标为机密的值是你账户的钥匙。千万不要分享给别人,也不要粘贴到聊天里。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config,把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"browserbase": {
"command": "npx",
"args": [
"-y",
"@mindstone/mcp-server-browserbase"
],
"env": {
"BROWSERBASE_API_KEY": "your-value",
"BROWSERBASE_REQUEST_TIMEOUT_MS": "your-value",
"MCP_WORKSPACE_PATH": "your-value"
}
}
}
}在终端中运行这条命令
claude mcp add browserbase --env BROWSERBASE_API_KEY=your-value --env BROWSERBASE_REQUEST_TIMEOUT_MS=your-value --env MCP_WORKSPACE_PATH=your-value -- npx -y @mindstone/mcp-server-browserbase运行前请先填写空着的值。
粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"browserbase": {
"command": "npx",
"args": [
"-y",
"@mindstone/mcp-server-browserbase"
],
"env": {
"BROWSERBASE_API_KEY": "your-value",
"BROWSERBASE_REQUEST_TIMEOUT_MS": "your-value",
"MCP_WORKSPACE_PATH": "your-value"
}
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"browserbase": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@mindstone/mcp-server-browserbase"
],
"env": {
"BROWSERBASE_API_KEY": "your-value",
"BROWSERBASE_REQUEST_TIMEOUT_MS": "your-value",
"MCP_WORKSPACE_PATH": "your-value"
}
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
BROWSERBASE_REQUEST_TIMEOUT_MS envMCP_WORKSPACE_PATH env这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月6日
注册名称已通过 GitHub 账户 mindstone 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
你需要填写:BROWSERBASE_API_KEY
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(0.1.1)于3个月前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。
同一分类中有:Browser Use, Chrome DevTools MCP, MCP Server Browser
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
MCP server for Chrome DevTools
MCP server for browser use access
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
Cloud browser automation via the Browserbase API: sessions with live-view debugging, AI web agents, persistent contexts, downloads, fetch/search, and serverless functions.
*Best for MCP hosts that want a local browser-automation connector covering the entire Browserbase API — from "open a browser and show me" to "run an agent that extracts pricing data".*
BROWSERBASE_API_KEY)STATUS.jsonBrowserbase's API spans everything from raw CDP sessions to fully-managed AI agents. This package exposes all of it to local stdio MCP hosts as a single connector with deliberate guardrails: session billing is called out in tool descriptions, destructive actions are annotated, external web content is always wrapped as untrusted data, and long-running work is polled (Browserbase has no webhooks).
"Extract the pricing tiers from example.com and give them to me as JSON."
Tools the host calls:
create_agent_run — starts a run with the task and a result_schema for the JSON shape.wait_for_agent_run — polls until the run completes and returns the structured result.get_agent_run_messages — (optional) shows what the agent did, step by step.Response (trimmed):
{
"ok": true, "runId": "r0a0b0c0-…", "status": "COMPLETED", "result": { "plans": [ { "name": "…", "price": 0 } ] }, "waited_seconds": 47 }
After clicking the button, your host will prompt you to fill: BROWSERBASE_API_KEY.
Manual config for Claude Desktop / Claude Code / Goose / Continue.dev (Browserbase)
{
"mcpServers": { "Browserbase": { "command": "npx", "args": [ "-y", "@mindstone/mcp-server-browserbase" ], "env": { "BROWSERBASE_API_KEY": "" } } } }
cd <path-to-repo>/connectors/browserbase
npm install npm run build
npx -y @mindstone/mcp-server-browserbase
node dist/index.js
| Variable | Required | Description |
|---|---|---|
BROWSERBASE_API_KEY | Yes | Browserbase API key. Get one at browserbase.com/settings |
BROWSERBASE_REQUEST_TIMEOUT_MS | No | Per-request timeout for Browserbase API calls in milliseconds (default 30000) |
MCP_WORKSPACE_PATH | No | Workspace directory for local files uploaded by upload_session_file, upload_extension, and upload_certificate. File reads are sandboxed to this directory; defaults to the system temp directory. |
When running inside a host that provides the credential bridge, configure_browserbase_api_key persists the key through the host instead of the environment.
{
"mcpServers": { "browserbase": { "command": "npx", "args": ["-y", "@mindstone/mcp-server-browserbase"], "env": { "BROWSERBASE_API_KEY": "your-api-key" } } } }
Every create_session starts a billable browser (per-minute, 1-minute minimum). The create_session and end_session descriptions tell the agent this explicitly, and end_session should be called as soon as automation finishes. Exceeding the project's concurrency limit returns 429 with a retry-after window that the error resolution surfaces verbatim.
get_session returns a credentialed connectUrl and get_session_debug_urls returns debugger/live-view URLs. Anyone with these URLs can watch AND drive the browser — they are for sharing with the session's owner, never publicly.
Anything authored outside your account — fetched page content, search results, page titles in debug URLs, agent run tasks/results/messages, CDP log bodies, replay playlists, download filenames, function output, error details — is wrapped in envelopes with close-tag breakout escaping before it reaches the model. CDP log rawBody values are additionally truncated beyond ~4KB with a truncation note.
upload_session_file, upload_extension, and upload_certificate accept local file_paths. Reads are constrained to MCP_WORKSPACE_PATH (or the system temp directory when unset) using canonical-prefix containment: .. traversal, absolute paths outside the sandbox, and symlinks that escape it are rejected before any disk read. The read is open-then-validate: the connector opens a descriptor, confirms the opened inode is the file the sandbox approved, reads through that descriptor, and re-checks it after the read — a file swapped or modified mid-read fails closed.
fetch_url (and the URLs embedded in search results you might fetch next) are guarded by a public-URL check: only http:///https:// schemes, and localhost, loopback, link-local, and private IP ranges are rejected before any request is sent.
delete_context, delete_agent, delete_download, delete_extension, delete_certificate, end_session, stop_agent_run, invoke_function, and all create_*/upload_* tools carry destructiveHint: true. Deletes are permanent; hosts should require explicit user confirmation.
GET /sessions/{id}/recording (rrweb event dump) is deprecated by Browserbase and returns a deprecation 404 — it is deliberately not exposed. Use get_session_replays / get_session_replay_playlist (watchable) or the recording-download tools (MP4 files) instead.wait_for_agent_run / get_agent_run), recording downloads (get_session_recording_downloads), and function invocations (get_function_invocation) are all polled.只显示了开头部分;其余内容请在仓库中阅读