Browser Use
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
作者による説明(英語の原文)
W3C WebMCP toolkit for Chrome 146. Expose any website as a structured MCP server for AI agents.
この3つを1つのスコアにまとめることはしません。意見、測定、人気はそれぞれ別のものだからです。
このサーバーのやさしい解説はまだ書いていません。以下はレジストリから読み取った事実です。
使っているAIアプリを選んで、設定をコピーしてください。
作者はレジストリにインストール情報を載せていません。下にある作者のREADMEをご覧ください。
パソコンで動くサーバーです。インストールすると、ツール一覧がAIアプリに表示されます。
誰でも見られる情報をチェックしており、コードは読んでいません。当てはまらない項目は数に入れていません。 · 2026年11月7日にチェック
レジストリ名は、リポジトリを所有するGitHubアカウント up2itnow0822 で確認済みです
このリポジトリはまだGitHubから読み取っていません
このリポジトリはまだGitHubから読み取っていません
これはパソコンで動くため、インターネット上のエンドポイントはありません
パソコンで動き、代わりにローカルの設定を使います
パソコンで動くサーバーです。ツール一覧を読み取れるのは、インターネット上のエンドポイントがある場合だけです
レジストリでこのカテゴリにある有効なすべてのサーバーと比較しています。スターは、公開者がリポジトリの所有者である場合だけ数えます。
当てはまる基本チェック1項目のうち1項目に合格しました。コードは読んでいません。キーやトークンを追加する前に、サーバーがどんなアクセスを求めているかを確認し、必要な権限だけを持つキーを使ってください。
レジストリには必須のキーは載っていません。ただし、最初に使うときにサインインを求めるサーバーもあります。
あなたのパソコンで動きます。必要なときにAIアプリが起動し、データは作者のサーバーを通りません(サーバー自体がオンラインサービスを呼び出す場合を除きます)。
GitHubリポジトリの最終更新は–、最新バージョン(0.5.0)がレジストリに届いたのは8 か月前です。
Claude Desktop、Claude Code、Cursor、VS CodeなどのMCP対応アプリで使えます。
同じカテゴリには:Browser Use, Chrome DevTools MCP, MCP Server Browser
同じカテゴリで、ほかの作者が作ったサーバー。スターの多い順。
カードの見方:★ GitHubスターi3つの柱:レビュー · チェック · 利用状況i更新=最新バージョンをレジストリに公開した日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
作者のリポジトリから書かれたとおりにコピーしたものです。PopNx MCPは内容を確認・推奨していません。HTML、画像、スクリプトは取り除いています。リンクはこのサイトの外へ移動します。
TypeScript tools for browser agents, with explicit authorization, bounded execution, React lifecycle support and a non-custodial transaction reference app.
0.6.0 is a breaking security release candidate. Existing registry releases do not contain these repairs. Read the migration guide. The candidate is unpublished; this project provides neither custody nor a guarantee against attacks.
From a repository checkout, with Node 22.22.2+ (22 LTS) or 24.15+ (24 LTS) and npm:
npm ci --ignore-scripts
npm run build npm run studio:dev
Open Vite's localhost URL. Practice mode uses invented balances, prices and receipts; it never signs or broadcasts. Live wallet mode uses an injected EIP-1193 wallet on Base chain 8453. The candidate supports ETH, native Base USDC and WETH transfers, and USDC/WETH exact-input swaps through the configured Uniswap V3 router and fee tier 500. Agents read state and propose actions; humans separately review each exact approval or transaction and authorize it in their external wallet.
Proposals bind account, chain, saved policy version, exact amount, minimum output and a 120-second expiry. Slippage is capped at 1%. Per-proposal limits do not cap aggregate spending. No key/seed input, arbitrary agent calldata, unlimited approval or agent signing tool is provided. Saved policy survives mode changes; memory-only activity clears on reload or mode change.
A submitted hash is pending. Matching transaction fields, receipt and token events must be checked before confirmation. Ambiguous submissions and RPC failures remain pending. Mined inclusion is not irreversible finality; inspect pending wallet activity before retrying. See the threat model and release runbook.
Install the reviewed package after publication, or build this branch locally. Public subpaths provide CommonJS, ES modules and TypeScript declarations.
import { createKit, defineTool } from "webmcp-sdk";
const kit = createKit({ prefix: "demo" }); kit.register( defineTool( "hello", "Return a greeting for the supplied name.", { type: "object", properties: { name: { type: "string", maxLength: 100 } }, required: ["name"], additionalProperties: false, }, ({ name }) => ({ message: Hello, ${name}! }), ), ); console.log(await kit.invoke("demo.hello", { name: "Bill" })); // { message: 'Hello, Bill!' } await kit.ready(); // Later: kit.unregisterAll();
Direct invocation works without a native browser API. The default adapter uses document.modelContext, execute, asynchronous registration and registration AbortSignals. Await ready() to observe registration errors. Native WebMCP is experimental and browser/origin dependent. Use the canonical browser example to distinguish local execution from native readiness. A legacy navigator adapter requires explicit browserAdapter: 'legacy' and separate compatibility testing. Annotations are advisory metadata, not authorization.
Inputs are finite own-data JSON snapshots, recursively frozen before asynchronous hooks. Dangerous keys, cycles, arbitrary prototypes, accessors and invalid schemas reject. Validation works without unsafe-eval; supported schema restrictions are intentional compatibility limits.
Defaults: 10 concurrent executions, 100 queued/admitted waits, 30-second deadlines, 65,536 input bytes, 1,000 nodes and depth 10. Handlers receive an optional second { signal } argument. Propagate cancellation and check current authority at the operation boundary. A deadline cannot preempt synchronous JavaScript or force an uncooperative promise to stop; unresolved work retains its capacity/lease until it settles.
import { withConfirmation, withSecurity } from "webmcp-sdk/security";
// reviewedTool, showReviewDialog and writeAuditEvent are application-owned. const secured = withSecurity( withConfirmation( reviewedTool, "Approve these reviewed terms?", showReviewDialog, ), { rateLimit: { maxInvocations: 10, windowMs: 60_000 }, audit: true, onAudit: writeAuditEvent, }, ); kit.register(secured);
withConfirmation requires its explicit provider to return literal true. kit.requestUserInteraction independently requires kit configuration confirmationProvider. Financial review must bind immutable terms; a generic boolean dialog is insufficient by itself. audit: true needs a reliable synchronous sink. Security audit omits input by default and never logs raw handler results/errors; explicitly allowed fields receive credential-key redaction. General kit callbacks/logs can contain input and are not a redacted audit store. blockedPatterns are bounded literal substrings, not a comprehensive prompt-injection filter.
Use WebMCPProvider from webmcp-sdk/react to scope session policy. Tool hooks publish committed handlers/metadata, revoke old registrations when callback authority changes, and remove only their owned registrations. Config-owned providers reject changed configuration; remount to replace policy. Supplied kits may be replaced. Stable callbacks avoid unnecessary revocation but must retain all account/policy dependencies.
registerDeclarativeTools(kit, root, { handlers }) registers annotated forms through an explicit trusted handler map and returns cleanup. HTML cannot define executable handlers. See declarative tools for supported fields and authorization responsibilities.
x402 payments: Paused for security maintenance. Do not use in production.
webmcp-sdk/middleware/express serves discovery metadata, not an execution transport or authentication system. Catalogs deny by default: supply an authorizer returning literal true, or deliberately choose publicManifest: true. Catalog responses use private, no-store. Actual operations need separate authentication and authorization.
Python discovery is a separate source distribution; npm does not install Python imports. No stdio server, webmcp-sdk serve, ProxyRelay or AuditLog class is supplied.
npm run verify:release
The release record distinguishes historical evidence from freshly reconstructed checks. The complete gate covers formatting, types, SDK/Studio tests and builds, packed consumers, native browser/CSP, deterministic wallet flows, accessibility/mobile checks, Python tests and dependency audit. Python installation has a separate wheel-import check. No automated check publishes, deploys or sends real funds. See SECURITY.md for reporting. MIT license.