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)
Developer fixes, free skills, browser tools and scoped technical services over MCP.
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。
必要なもの
どのアプリを使っていますか?
上でアプリを選んでください。選んだアプリを覚えておくので、次からはボタンに「…にインストール」と表示されます。
ウェブ、デスクトップ、モバイルのClaudeで使えます:
https://knownfix-backend-28.b-hash88.deno.net/mcpClaude Desktop → Settings → Developer → Edit Config を開き、claude_desktop_config.json にこれを貼り付けます
{
"mcpServers": {
"knownfix": {
"command": "npx",
"args": [
"-y",
"knownfix",
"bridge"
]
}
}
}ターミナルでこれを実行します
claude mcp add --transport http knownfix https://knownfix-backend-28.b-hash88.deno.net/mcpclaude mcp add knownfix -- npx -y knownfix bridgeプロジェクトの .cursor/mcp.json、または ~/.cursor/mcp.json に貼り付けます
{
"mcpServers": {
"knownfix": {
"url": "https://knownfix-backend-28.b-hash88.deno.net/mcp"
}
}
}プロジェクトの .cursor/mcp.json、または ~/.cursor/mcp.json に貼り付けます
{
"mcpServers": {
"knownfix": {
"command": "npx",
"args": [
"-y",
"knownfix",
"bridge"
]
}
}
}ブラウザがアプリを開いてよいか確認し、そのあとアプリが設定を表示して確認を求めます。キーの値は空欄のままなので、自分で入力してください。
プロジェクトの .vscode/mcp.json に貼り付けます
{
"servers": {
"knownfix": {
"type": "http",
"url": "https://knownfix-backend-28.b-hash88.deno.net/mcp"
}
}
}プロジェクトの .vscode/mcp.json に貼り付けます
{
"servers": {
"knownfix": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"knownfix",
"bridge"
]
}
}
}ブラウザがアプリを開いてよいか確認し、そのあとアプリが設定を表示して確認を求めます。キーの値は空欄のままなので、自分で入力してください。
ChatGPTがつなげるのは、インターネット上のサーバーだけです:
https://knownfix-backend-28.b-hash88.deno.net/mcpメニューの名前は、プランや時期によって少し違うことがあります。
選んだアプリは、この端末のCookieに保存されます。アカウントは不要です。
この設定は、作者のレジストリ登録情報から作成しています。READMEと違う場合は、READMEに従ってください。
このサーバーのツール一覧はまだ読み取っていません。エンドポイントは順番にチェックしています。
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 b-hash88 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
まだチェックしていません。エンドポイントは順番にチェックしています
まだチェックしていません。エンドポイントは順番にチェックしています
まだチェックしていません。エンドポイントをチェックするときにツール一覧を読み取ります
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。
どちらでも使えます。パソコンにインストールしても、クラウド版につないでもかまいません。
GitHubリポジトリの最終更新は–、最新バージョン(0.6.1)がレジストリに届いたのは18 日前です。
Claude Desktop、Claude Code、Cursor、VS CodeなどのMCP対応アプリで使えます。URLでつながるので、リモートMCPサーバーに対応したアプリ(開発者モードのChatGPTなど)でも使えます。
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。
KnownFix is an agent-facing catalog of development-error fixes. The current overview reports inventory and evidence labels; fourteen fixes are available in full for free. Search is free; paid fixes start at $0.05 USDC through Base Pay or a signed exact-ETH equivalent on Base mainnet.
KnownFix offers ten scoped services, including website audits, a homepage build, code and release reviews, a Webhook Reliability Review ($179 USDC / 3 business days), and a SaaS Security Exposure Review ($299 USDC / 5 business days). Verified Repo Fix is $149 USDC / 3 business days from scope acceptance and a complete reproducible brief, with payment only on verified delivery. Microsite Launch is $299 USDC / 5 business days after scope approval, approved content/assets and manual payment before work. SaaS security requires written ownership/testing permission and an agreed target/method/window allowlist before payment or testing; its five-day target begins after approval, payment and required access. These three services use manual scope/payment arrangements, not automated checkout. Other review targets begin after payment and a complete in-scope brief. The service catalog is authoritative.
Three free browser-local tools complement six free workflow skills: MCP Connection Kit, MCP Response Inspector, and Webhook Retry Planner. The planner compares capped backoff and full-jitter timing against a time budget. It sends no webhooks and saves no settings. Optional analytics records fixed action labels only; those events are not unique users or successful deliveries.
Website analytics uses GA4 G-PPPE9TYRM2 (property 559407250). The earlier G-781DC76YQN is the Fourthwall shop stream, not the website reporting source. Never merge shop visits or MCP request counters into website visitor totals. The website property reports in America/Vancouver; a UTC test can fall on the previous reporting day. Consent is opt-in, advertising stays denied, and QA events are not customer activity. Report processing is separate from transport.
tests/analytics-property.mjs checks every HTML page and marked action, failing on missing tracking rather than silently skipping new pages. Four deliberate exclusions stay untracked: the ownership-verification file, fictional Patchnote demo and two sample reports. tests/analytics-consent-lifecycle.mjs covers every tracked wrapper. tests/analytics-coverage-browser.mjs loads all tracked pages and exercises homepage actions with Google and backend writes blocked; it uses the installed-Playwright recipe below. tests/tool-guides-browser.mjs verifies the three tools' check, copy, download, calculate, reset and guide actions. The opt-in tests/tool-guides-live-analytics.mjs creates synthetic QA events in temporary profiles only. HTTP204 is transport evidence, not proof of GA4 report receipt; verify the matching main-site property separately. Never alter an owner's saved consent or use a live checkout to test analytics.
Three companion guides live under guides/, with nine fixed synthetic JSON examples, linked from the tools, homepage and free-skills hub. Regenerate this content with node scripts/sync-tool-guides.mjs. The MCP-kit and retry-planner generators also run this sync. node tests/tool-guides.mjs checks reproducible outcomes, links, downloads, analytics boundaries and generator idempotence. tests/tool-guides-browser.mjs uses the same installed-Playwright environment recipe below; KNOWNFIX_BROWSER_CHANNEL=msedge selects an installed Edge. Guide events use guide_action with fixed open-tool, download-example and related-guide contexts. Tool examples reuse the respective tool action event; no pasted input or entered settings enter these events. All require consent.
Run node tests/retry-planner.mjs and node tests/service-comparison.mjs for the new tool and catalog consistency. The existing Gitea workflow includes both. tests/retry-planner-browser.mjs uses the existing KNOWNFIX_PLAYWRIGHT_PATH recipe to resolve an installed Playwright package without downloading one. Set PLAYWRIGHT_BROWSERS_PATH to the installed browser cache when needed and KNOWNFIX_SCREENSHOT_DIR to retain desktop/mobile screenshots. Tests start and close a temporary localhost server and block real backend writes.
Regenerate the planner with node scripts/sync-retry-planner.mjs, then run node scripts/sync-site-improvements.mjs to refresh catalog comparison and links. The service sync also refreshes that comparison after catalog changes.
Use the client-specific connection kit for Claude Code, Codex, Cursor, VS Code or the Claude Desktop stdio bridge. It includes supported install links and a free first-use example. Configuration, tool discovery and successful workflow retrieval are distinct milestones.
Six standard skills/ /SKILL.md directories contain the complete free workflows, synthetic exercises and acceptance criteria. Review them, then use npx skills add b-hash88/knownfix to choose a client and scope interactively, or copy reviewed directories using the client's documented skill installation.
The toolkit also provides dependency-free local commands:
node toolkit/cli.mjs inspect-response toolkit/examples/response.json
node toolkit/cli.mjs retry-plan toolkit/examples/retry.json
Node.js 18+, file/stdin JSON, 200,000-byte cap. No network or telemetry during these commands. Envelope structure is not tool success; modeled retries do not send webhooks. Run node scripts/sync-agent-toolkit.mjs after changing skill or browser-core sources, and node tests/agent-toolkit.mjs to verify parity.
KnownFix is listed as io.github.b-hash88/knownfix. Its remote Streamable HTTP MCP endpoint is:
https://knownfix-backend-28.b-hash88.deno.net/mcp
The live registry exposes search, catalog, signed offers, fix and skill delivery, endpoint audits, submissions, fix requests, professional review ordering and retrieval, and a read-only apparel catalog through list_merch. npx knownfix tools prints the current list instead of relying on copied docs. The branded Smithery directory name is knownfix/knownfix.
For the current npm Trusted Publishing failure, search the exact CLI signature:
npx knownfix search "npm ERR! code EOTP"
KnownFix also matches This operation requires a one-time password from your authenticator. and the npm --otp= wording. The public diagnostic page cites npm's current CLI source; the npm Publishing Recovery Pack contains the six-fix decision tree and verification commands.
Free fixes return immediately. A paid search_fixes match includes a free diagnosis preview, confidence and compatibility, dollar price, a private signed offer, wallet/payment URI, and one recommended next action. Each one-hour offer is bound to one product, one currency, and one payment intent:
search_fixes, or request one directly with MCP get_offer or POST /offer.priceWei for ETH on chain 8453.paymentTx plus paymentOffer, or the corresponding HTTP headers. For USDC, paymentTx is the returned ERC-4337 UserOperation hash; for ETH, it is the mined transaction hash.The payment proof and offer are atomically single-use in Deno KV. A public chain hash alone is not accepted as a purchase credential. The store also carries five crawlable recovery packs for npm, GitHub Actions, MCP, Windows agent shells, and Base payment verification.
verified-in-production means reproduced and fixed in a real session; documented means sourced but not independently reproduced. An unmatched search returns empty. The books distinguish raw requests, MCP polling, real tool calls, deliveries, paid intent, and settled sales.
KnownFix is AI-operated on human-owned hosting, deployment, and treasury infrastructure. No buyer account or API key is required.
Run the dependency-free production suite with Node 20 or newer:
node tests/live-e2e.mjs
The suite crawls every canonical URL and checks the current free/paid page pairs, product packs and service catalog, HTTP and MCP search/delivery/offer flows, distinct UserOperation and transaction proofs, anonymous stats access controls, and security headers. Requests carry x-operator: 1 so delivery and purchase-intent tests do not contaminate those conversion counters. Set KNOWNFIX_SERVICE_ORDER_TARGET to an authorized public URL only when intentionally testing creation of one private operator order; routine runs do not fill the order queue.
agents.html is the human-readable MCP start page; notes-for-agents.md is its portable counterpart. Start with get_overview / knownfix://overview / backend GET /overview. get_free_skill({id}) supports tool-only clients; resource-capable clients can read each knownfix://free-skill/ directly. There are three local browser tools, three guides, six full free workflows, eleven premium workflow outlines and ten scoped services. Manual approval and payment-on-delivery services never use upfront automated orders.
After other storefront generators, run node scripts/sync-site-quality.mjs. The guide and comparison generators invoke this finalizer too. It preserves prices/scopes, aligns payment metadata, maintains fix landmarks/navigation, and applies current-consent/own-property analytics guards. premium-skills.json contains only allowlisted fields from the public backend /skills response, not paid bodies. Refresh that snapshot only after checking current public terms. node tests/site-quality.mjs protects consistency and generator idempotence. scripts/audit-site.mjs uses the same explicitly configured installed Playwright runtime as the browser suites to inspect all public HTML and local links.