/ Annuaire / Playground / Browser Tools
● Communauté AgentDeskAI ⚡ Instantané

Browser Tools

par AgentDeskAI · AgentDeskAI/browser-tools-mcp

Read live console logs, network errors, DOM, and screenshots from YOUR real Chrome — debugging context straight into your agent.

AgentDeskAI's browser-tools MCP ships a Chrome extension + local bridge + MCP server. Your agent sees what YOU see — logged-in sessions, extensions, the real DOM and console. Best for debugging sessions where reproducing in a clean browser is the problem.

Pourquoi l'utiliser

Fonctionnalités clés

Démo en direct

Aperçu en pratique

browser-tools.replay ▶ prêt
0/0

Installer

Choisissez votre client

~/Library/Application Support/Claude/claude_desktop_config.json  · Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "browser-tools": {
      "command": "npx",
      "args": [
        "-y",
        "@agentdeskai/browser-tools-mcp@latest"
      ]
    }
  }
}

Ouvrez Claude Desktop → Settings → Developer → Edit Config. Redémarrez après avoir enregistré.

~/.cursor/mcp.json · .cursor/mcp.json
{
  "mcpServers": {
    "browser-tools": {
      "command": "npx",
      "args": [
        "-y",
        "@agentdeskai/browser-tools-mcp@latest"
      ]
    }
  }
}

Cursor utilise le même schéma mcpServers que Claude Desktop. La config projet l'emporte sur la globale.

VS Code → Cline → MCP Servers → Edit
{
  "mcpServers": {
    "browser-tools": {
      "command": "npx",
      "args": [
        "-y",
        "@agentdeskai/browser-tools-mcp@latest"
      ]
    }
  }
}

Cliquez sur l'icône MCP Servers dans la barre latérale Cline, puis "Edit Configuration".

~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "browser-tools": {
      "command": "npx",
      "args": [
        "-y",
        "@agentdeskai/browser-tools-mcp@latest"
      ]
    }
  }
}

Même format que Claude Desktop. Redémarrez Windsurf pour appliquer.

~/.continue/config.json
{
  "mcpServers": [
    {
      "name": "browser-tools",
      "command": "npx",
      "args": [
        "-y",
        "@agentdeskai/browser-tools-mcp@latest"
      ]
    }
  ]
}

Continue utilise un tableau d'objets serveur plutôt qu'une map.

~/.config/zed/settings.json
{
  "context_servers": {
    "browser-tools": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "@agentdeskai/browser-tools-mcp@latest"
        ]
      }
    }
  }
}

Ajoutez dans context_servers. Zed recharge à chaud à la sauvegarde.

claude mcp add browser-tools -- npx -y @agentdeskai/browser-tools-mcp@latest

Une seule ligne. Vérifiez avec claude mcp list. Supprimez avec claude mcp remove.

Cas d'usage

Usages concrets : Browser Tools

Debug a live front-end bug using real console + network context

👤 Frontend engineers ⏱ ~15 min intermediate

Quand l'utiliser : A bug only reproduces in your actual session (logged in, specific data). You want the agent to read the console and network tab.

Prérequis
  • Chrome extension installed + bridge server running — Install extension from releases page, then npx @agentdeskai/browser-tools-server@latest in a terminal
  • MCP server wired to your clientnpx -y @agentdeskai/browser-tools-mcp@latest
Déroulement
  1. Reproduce the bug with DevTools open
    Open the affected page in Chrome with the extension active. Trigger the bug. Don't close the tab.✓ Copié
    → Bug visible, extension icon shows it's capturing
  2. Pull context into the agent
    Get the last 50 console messages and any failed network requests. Summarize what the page is yelling about.✓ Copié
    → Specific errors surfaced with stack traces
  3. Diagnose and propose fix
    Based on the errors, point at the file/line likely responsible. If 404s, identify the bad URL; if JS errors, the root cause.✓ Copié
    → Concrete hypothesis with evidence

Résultat : A debugging loop where the agent has the same context as you opening DevTools — faster than reading a screenshot description.

Pièges
  • Bridge server not running, MCP silently returns empty data — Always check curl localhost:3025/identity responds first; the bridge must be running before the MCP call
  • Capturing from the wrong tab — Extension captures the active tab. Click into the right tab before asking the agent
Combiner avec : sentry

Run a Lighthouse-style audit on the page you're looking at

👤 Frontend / SEO / a11y engineers ⏱ ~10 min beginner

Quand l'utiliser : You want to audit a page as it actually renders in your session (not a clean crawler).

Déroulement
  1. Open the target page
    Navigate to <URL> in Chrome with the extension active.✓ Copié
    → Page loaded
  2. Run the audit mode
    Run the audit tool — perf, a11y, SEO, best practices. Summarize top 5 issues with fix suggestions.✓ Copié
    → Prioritized issue list
  3. Focus in
    For the accessibility failures, list exact selectors and the WCAG rule violated.✓ Copié
    → Actionable remediation steps

Résultat : A full audit without leaving your browser, scoped to real user state.

Pièges
  • Audits for auth-gated pages fail because the tool re-navigates — Use the debugger mode's element/console tools on the already-loaded page instead of the full audit

Inspect and explain a specific element

👤 Frontend devs, designers ⏱ ~5 min beginner

Quand l'utiliser : Something looks off in one component. You want a quick 'what CSS is applied and why' without digging.

Déroulement
  1. Select the element in DevTools
    Open DevTools, click the element inspector, select the offending element.✓ Copié
    → Element highlighted in Elements panel
  2. Ask the agent to inspect
    Get the selected element's computed styles, inherited rules, and bounding box. Explain why it's rendering the way it is.✓ Copié
    → Explanation grounded in actual CSS cascade
  3. Propose a fix
    Suggest the smallest CSS change to achieve [desired visual]. Show the before/after rule.✓ Copié
    → Minimal diff suggestion

Résultat : CSS debugging at conversation speed.

Pièges
  • Element selection doesn't persist — agent sees a different one — Re-select the element right before asking; don't let the DevTools panel change focus between messages

Combinaisons

Associez-le à d'autres MCPs pour un effet X10

browser-tools + sentry

Reproduce a Sentry error in YOUR browser and pull live console context

Sentry issue WEB-3a91 says an XHR to /api/checkout returns 500. Navigate to the page in my Chrome, reproduce, and capture the console + network.✓ Copié
browser-tools + playwright

browser-tools for live debugging, Playwright for repeatable reproductions afterward

I found the bug in my live Chrome via browser-tools. Now write a Playwright test that reproduces the exact click sequence.✓ Copié

Outils

Ce que ce MCP expose

OutilEntréesQuand appelerCoût
takeScreenshot Visual capture of current state free
getConsoleLogs / getConsoleErrors Debug any JS error free
getNetworkLogs / getNetworkErrors Diagnose failed API calls free
getSelectedElement User has selected an element in DevTools free
runAccessibilityAudit A11y review of current page free
runPerformanceAudit Perf audit of current page free
runSEOAudit SEO review free
runBestPracticesAudit General code quality review free
runNextJSAudit You're on Next.js free
wipeLogs Clear between investigation sessions free

Coût et limites

Coût d'exécution

Quota d'API
None — all local
Tokens par appel
Log dumps can be big (5k+ tokens). Use wipeLogs between sessions.
Monétaire
Free, open source
Astuce
Wipe logs before each investigation — avoids pulling in unrelated noise

Sécurité

Permissions, secrets, portée

Stockage des identifiants : None — runs against your existing Chrome session
Sortie de données : Only to localhost:3025 (the bridge). Nothing leaves your machine except via whatever the agent/MCP client does.

Dépannage

Erreurs courantes et correctifs

All tools return empty / 'not connected'

The bridge server isn't running. Start it: npx @agentdeskai/browser-tools-server@latest in a dedicated terminal.

Vérifier : curl http://localhost:3025/identity
Extension icon grey / inactive

Click the extension icon, confirm it says 'Connected'. If not, reload the tab.

Audits fail with timeout

Some audits need extra time on heavy pages. Close other tabs to free CPU; re-run.

Wrong tab captured

The extension captures the tab that was last activated. Click into the right tab immediately before calling a tool.

Alternatives

Browser Tools vs autres

AlternativeQuand l'utiliserCompromis
chrome-devtools MCPYou want direct CDP access without a browser extensionNo extension needed but you lose real-user-session context
Playwright MCPYou need repeatable automation, not live debuggingClean browser — can't see your logged-in session

Plus

Ressources

📖 Lire le README officiel sur GitHub

🐙 Voir les issues ouvertes

🔍 Parcourir les 400+ serveurs MCP et Skills