Crawlio Browser

Documentation | API Reference | Chrome Extension

MCP server that gives AI full control of a live Chrome browser via CDP. 145 tools with framework-aware intelligence, typed evidence infrastructure, tracking pixel analysis, technographic fingerprinting, SEO auditing, and confidence-tracked findings — captures what static crawlers can't see.

When to use Crawlio Browser

Use Crawlio Browser when your AI needs to interact with a real browser — SPAs, authenticated pages, dynamic content, JS-rendered frameworks. Unlike headless browser tools, Crawlio Browser connects to your actual Chrome via a lightweight extension, giving the AI access to your logged-in sessions, cookies, and full browser state.

Crawlio Browser vs headless browser tools: Headless tools launch a separate browser process. Crawlio Browser connects to your existing Chrome — no separate browser, no login flows, full access to your tabs and sessions.

[!WARNING] This is the trade-off, stated plainly: connecting your own Chrome is the feature, and it means an AI agent can act as you on every site you are logged into. It attaches Chrome's debugger to the tab you connect, so it can read cookies, storage, and page content for that session.

Review what you connect it to. Prefer a dedicated Chrome profile for agent work. Nothing is captured until you connect a tab, and you can see the exact tool surface before configuring anything by running npx crawlio-browser tools. Sites can opt out with <meta name="crawlio-agent" content="disable">, which the extension honors.

Quick Start

1. Install the Chrome Extension
2. Run the init wizard:

   npx crawlio-browser init
   

That's it. Auto-detects and configures 14 MCP clients: Claude Code, Cursor, VS Code, Codex, Gemini CLI, Claude Desktop, ChatGPT Desktop, Windsurf, Cline, Zed, Goose, OpenCode, MCPorter, and Cline CLI.

Init wizard options

npx crawlio-browser init              # Default: code mode, stdio transport
npx crawlio-browser init --full       # Full mode (every tool exposed individually)
npx crawlio-browser init --portal     # Portal mode (persistent HTTP server)
npx crawlio-browser init --cloudflare # Add Cloudflare MCP (89 tools, no wrangler)
npx crawlio-browser init --dry-run    # Show what would happen
npx crawlio-browser init --yes        # Skip prompts (CI / scripted installs)
npx crawlio-browser init -a claude    # Target specific MCP client

Inspecting what it exposes

npx crawlio-browser tools          # What code mode exposes (the default)
npx crawlio-browser tools --full   # Every tool, individually
npx crawlio-browser tools --json   # Machine-readable, for diffing across versions
npx crawlio-browser doctor         # Bridge, portal, native host, client configs
npx crawlio-browser --help         # All commands and options
npx crawlio-browser --version      # Version only

Both are read-only and run without a browser, an extension, or a network connection — you can see the whole surface before you configure any client. The numbers come from the same builders the server registers, so they cannot disagree with what your client receives.

Transport Modes

Portal mode is recommended for Claude Code — the server persists across context compaction and session restarts. On macOS, --portal installs a launchd agent for auto-start on login.

Manual setup (any client)

{ "mcpServers": { "crawlio-browser": { "command": "npx", "args": ["-y", "crawlio-browser"] } } }

{ "mcpServers": { "crawlio-browser": { "type": "http", "url": "http://127.0.0.1:3001/mcp" } } }

claude mcp add crawlio-browser -- npx -y crawlio-browser

{ "mcpServers": { "crawlio-browser": { "command": "npx", "args": ["-y", "crawlio-browser"] } } }

{ "mcpServers": { "crawlio-browser": { "command": "npx", "args": ["-y", "crawlio-browser"] } } }

{ "cline.mcpServers": { "crawlio-browser": { "command": "npx", "args": ["-y", "crawlio-browser"] } } }

How It Works

AI Client (stdio/http)  -->  MCP Server (Node.js)  -->  Chrome Extension (MV3)
                             crawlio-browser               WebSocket -> CDP

The MCP server communicates with the Chrome extension via WebSocket. The extension controls the browser through Chrome DevTools Protocol (CDP).

Capabilities

Framework-Aware Intelligence

Every execute call probes the browser for framework signatures and injects a shape-shifting smart object with framework-native accessors. React state, Vue reactivity, Next.js routing, Shopify cart data — 17 framework namespaces across 4 tiers, detected at runtime and rebuilt on every navigation. The AI doesn't query a generic DOM; it queries the framework's own data structures.

Evidence-Based Analysis

Method Mode adds higher-order methods and a typed evidence system on top of Code Mode. smart.extractPage() runs 7 parallel operations in a single call — page capture, performance metrics, security state, font detection, meta extraction, accessibility audit, and mobile-readiness check. Failed operations produce typed CoverageGap records instead of silent nulls. Findings created with smart.finding() get their confidence automatically adjusted when supporting data is missing. The result: structured, auditable research output with gap tracking and confidence propagation.

Session Recording & Replay

Record browser interactions as structured data, then compile them into reusable SKILL.md automations. 12 interaction tools are automatically intercepted during recording — clicks, typing, navigation, scrolling — each capturing args, result, timing, and page URL. One compileRecording() call converts the session into a deterministic automation script.

Robot Training

Capture human-guided browser demonstrations as replayable robot-training bundles. robot_training_start opens a fresh connected tab, starts CDP network capture and session recording, injects an event-driven state monitor, and robot_training_stop persists recording.json, network.json, response bodies, state snapshots, and flows.jsonl for API synthesis or background replay.

Auto-Settling & Actionability

Every mutative action (click, type, navigate, select_option) runs actionability checks before acting — polling visibility, dimensions, enabled state, and overlay detection. After the action, a progressive backoff settle delay ([0, 20, 100, 100, 500]ms) waits for DOM mutations to quiesce. The AI doesn't need manual sleep() calls between actions.

Architecture

A layered execution architecture where each layer absorbs a category of complexity that would otherwise fall on the model. The model sees three tools and a clean SDK. Everything beneath that surface is the runtime absorbing reality.

The layer worth understanding is the one that assembles itself. Detection runs against the live page on first use, and the smart object is built to match what that page turned out to be — smart.react.* exists only where React does. Nothing about the target is known at startup, so the surface is composed per tab rather than declared up front.

                     ┌───────────────────────────────────┐
                     │        AI Model (LLM)             │
                     │  Writes code, reads errors, loops  │
                     └───────────────┬───────────────────┘
                                     │  search, execute, connect_tab (+ 3 job tools)
                                     ▼
┌─────────────────────────────────────────────────────────────────┐
│                     Crawlio Browser runtime                      │
│                                                                  │
│  ┌────────────────────────────────────────────────────────────┐  │
│  │  METHOD MODE                                               │  │
│  │  Behavioral protocol + higher-order methods                │  │
│  │  scrollCapture · waitForIdle · extractPage · comparePages  │  │
│  │  detectTables · extractTable · waitForNetworkIdle ·        │  │
│  │  extractData                                               │  │
│  │                                                            │  │
│  │  ↳ Absorbs: behavioral variance, ad-hoc composition,      │  │
│  │    inconsistent output shapes, data extraction patterns    │  │
│  ├────────────────────────────────────────────────────────────┤  │
│  │  POLYMORPHIC CONTEXT                                       │  │
│  │  17 framework namespaces, injected Just-In-Time            │  │
│  │  react · vue · angular · nextjs · shopify · ...            │  │
│  │                                                            │  │
│  │  ↳ Absorbs: framework opacity, minified code,             │  │
│  │    devtools hook complexity                                │  │
│  ├────────────────────────────────────────────────────────────┤  │
│  │  ACTIONABILITY ENGINE                                      │  │
│  │  7 core smart methods with built-in resilience             │  │
│  │  click · type · navigate · waitFor · evaluate ·            │  │
│  │  snapshot · screenshot                                     │  │
│  │                                                            │  │
│  │  ↳ Absorbs: DOM timing, hydration delays, CSS animations, │  │
│  │    disabled states, overlapping elements                   │  │
│  ├────────────────────────────────────────────────────────────┤  │
│  │  TETHERED IPC BRIDGE                                       │  │
│  │  WebSocket ↔ Chrome extension, message queue,              │  │
│  │  heartbeat, auto-reconnect, stale detection                │  │
│  │                                                            │  │
│  │  ↳ Absorbs: connection drops, tab refreshes,              │  │
│  │    port conflicts, extension lifecycle                     │  │
│  ├────────────────────────────────────────────────────────────┤  │
│  │  COMMAND CHANNEL                                           │  │
│  │  bridge.send → CDP browser control via the extension       │  │
│  │  crawlio.*   → Crawlio HTTP endpoints                      │  │
│  │  178 searchable across both (145 browser + 33 HTTP)        │  │
│  └────────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
                                     │
                                     ▼
                     ┌───────────────────────────────────┐
                     │         Live Chrome Browser        │
                     │   Persistent session, real DOM,    │
                     │   framework runtime, user state    │
                     └───────────────────────────────────┘

What Each Layer Absorbs

Execution Lifecycle

1. Discovery — Model calls search("page capture performance") and gets documentation for relevant commands
2. Framework Detection — Runtime probes the live DOM, detects active frameworks, constructs polymorphic smart object with appropriate namespaces
3. Scope Assembly — Model's code is compiled into an async function with injected parameters: bridge (the browser command channel), crawlio (HTTP client), sleep, TIMEOUTS, smart (7 core + 18 higher-order methods + up to 17 framework namespaces), compileRecording
4. Execution — Method Mode methods compose the lower layers: extractPage() fires 7 parallel bridge.send() calls; click() runs the actionability engine; react.getVersion() evaluates framework-specific expressions
5. Error Recovery (Agentic REPL) — On failure, the browser stays in the exact state that produced the error. The model reads the structured error, adjusts, and calls execute again. Framework cache persists — no re-detection unless URL changed

Design Principles

1. Absorb complexity downward — Every category of difficulty (connection management, DOM timing, framework detection, multi-step composition) is handled by the layer best equipped for it. The model only encounters the clean interface at the top.
2. Shape the SDK to the target — The polymorphic context system detects what the page is and reshapes available methods to match. The model writes against a stable interface; the runtime adapts underneath.
3. Preserve state across cycles — The tethered architecture means the model can fail, learn, and retry against the same live environment — transforming error handling from "restart from scratch" into "adjust and continue."

How It Compares

Code Mode is Cloudflare's idea and a good one: present tools as a typed API and let the model write code against it, because models have seen far more code than tool calls. Crawlio Browser applies that pattern to a target it was not built for — a live browser holding your session. This is an MCP server, not an alternative to MCP.

The row that matters is error recovery. An isolate is disposable by design — that is what makes it safe to run untrusted code, and it means a failure discards the state that caused it. Here execution is tethered to a real tab, so when code fails the browser is still sitting in exactly the situation that broke it: same scroll position, same modal, same half-filled form. The model reads the error and runs again against that. Neither approach is better in general; they are answers to different problems, and the sandboxing guarantees Cloudflare gets from V8 isolates are genuinely stronger than what a local async sandbox provides.

Read the full architecture guide →

Two Modes

Code Mode (3 primary tools) — default

Collapses 145 tools into three high-level tools, plus get_job_result, list_jobs, and cancel_job for async execution — so tools/list reports

That makes the tools/list payload 83% smaller, measured by serializing the result each mode actually returns rather than estimated from a tool count. Code mode is not free: execute and search carry long descriptions, which is why the figure is 83% and not the ~95% a naive tools-only ratio suggests. Check it yourself — npx crawlio-browser tools --json prints both surfaces without connecting to anything.

// Navigate and screenshot
await bridge.send({ type: 'browser_navigate', url: 'https://example.com' }, 30000);
await sleep(2000);
const screenshot = await bridge.send({ type: 'take_screenshot' }, 10000);
return screenshot;

Full Mode (145 tools)

Every tool exposed directly to the LLM. Enable with --full:

npx crawlio-browser init --full

Smart Object

In Code Mode, the smart object provides framework-aware helpers with auto-waiting and actionability checks.

Core Methods

Higher-Order Methods

Typed Evidence

Methods for structured analysis findings with confidence propagation:

When a finding's dimension matches an active coverage gap, confidence is automatically capped:

Framework Namespaces

When a framework is detected, the smart object exposes framework-specific helpers:

Method Mode

Code Mode asks the model to write code. Method Mode gives that code a tested vocabulary:

would otherwise improvise.

await smart.extractTable(selector)   // not a hand-rolled scrape loop
await smart.scrollCapture()          // not a guessed scroll cadence

It is a domain layer over Code Mode, not a replacement: the tool surface does not change, the model still sees the same three primary tools, and the same 178-command catalog sits underneath. What changes is what happens inside execute.

The surface is assembled per page. Detection recognizes 64 frameworks and attaches up to 17 matching namespaces, so smart.react.* exists only where React does — the model never has to ask what the page is built with, or guess at hook paths and fiber internals to find out.

And because execution is tethered to a live tab rather than a disposable isolate, a failure leaves the browser in the exact state that produced it. The model reads the structured error and runs again against that state, rather than rebuilding the situation from scratch.

The Maturity Ladder

Architecture

┌────────────────────────────────────────────────────────────┐
│                      execute sandbox                       │
│                                                            │
│  ┌──────────────────────────────────────────────────────┐  │
│  │  Behavioral Protocol  (web-research skill)           │  │
│  │  Acquire → Normalize → Analyze                       │  │
│  ├──────────────────────────────────────────────────────┤  │
│  │  Evidence Infrastructure                             │  │
│  │  finding() · findings() · clearFindings()            │  │
│  │  Typed records · Coverage gaps · Confidence prop.    │  │
│  ├──────────────────────────────────────────────────────┤  │
│  │  Higher-Order Methods  [18]                          │  │
│  │  scrollCapture · waitForIdle · extractPage ·         │  │
│  │  comparePages · detectTables · extractTable ·        │  │
│  │  waitForNetworkIdle · extractData · detectSections · │  │
│  │  detectTechnologies · parseTrackingPixels · ...      │  │
│  ├──────────────────────────────────────────────────────┤  │
│  │  Smart Core  [7 methods]                             │  │
│  │  evaluate · click · type · navigate · waitFor ·      │  │
│  │  snapshot · rebuild                                  │  │
│  ├──────────────────────────────────────────────────────┤  │
│  │  Framework Namespaces  [up to 17, attached per page] │  │
│  │  react · vue · angular · nextjs · shopify · ...      │  │
│  ├──────────────────────────────────────────────────────┤  │
│  │  bridge.send()  — the browser command channel        │  │
│  └──────────────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────────────┘

Each layer up encodes more domain knowledge. bridge.send({ type: "capture_page" }) captures a page. smart.extractPage() captures a page AND runs performance metrics, security state, font detection, accessibility analysis, and mobile-readiness checks in parallel — seven operations, one call, graceful failure on supplementary data, typed gaps for anything that fails.

Evidence Infrastructure

Coverage Gaps — When supplementary operations in extractPage() fail, they don't silently return null. A typed gap is recorded with the dimension, reason, impact, and whether it reduces confidence on related findings:

// Example gap from a failed performance metrics call
{ dimension: "performance", reason: "CDP domain disabled", impact: "method-failed", reducesConfidence: true }

Tool-Enforced Findings — smart.finding() validates every field at the tool level. The model cannot produce a finding without meeting the schema — it either returns a valid Finding or gets a clear error. Findings accumulate across execute calls within a session via smart.findings().

Session Aggregation — Findings and coverage gaps persist across execute calls. A model can make findings across multiple calls, then retrieve the full set with smart.findings(). Reset with smart.clearFindings().

End-to-End Example: Competitive Audit

// 1. Extract and compare both sites (scaffold + gaps included)
const comparison = await smart.comparePages(
  'https://acme.com',
  'https://rival.com'
);

// 2. Make findings — confidence auto-adjusts based on data availability
smart.finding({
  claim: 'Rival loads 2.3x faster on Largest Contentful Paint',
  evidence: [
    `Acme LCP: ${comparison.siteA.performance?.webVitals?.lcp}ms`,
    `Rival LCP: ${comparison.siteB.performance?.webVitals?.lcp}ms`,
  ],
  sourceUrl: 'https://acme.com',
  confidence: 'high',
  method: 'comparePages + extractPage performance metrics',
  dimension: 'performance',  // if perf data failed, confidence caps to "medium"
});

smart.finding({
  claim: 'Acme has 12 images without alt text; Rival has 0',
  evidence: [
    `Acme imagesWithoutAlt: ${comparison.siteA.accessibility?.imagesWithoutAlt}`,
    `Rival imagesWithoutAlt: ${comparison.siteB.accessibility?.imagesWithoutAlt}`,
  ],
  sourceUrl: 'https://acme.com',
  confidence: 'high',
  method: 'comparePages + extractPage accessibility summary',
  dimension: 'accessibility',
});

// 3. Capture visual evidence
await smart.navigate('https://acme.com');
await smart.waitForIdle();
const acmeVisuals = await smart.scrollCapture({ maxSections: 5 });

// 4. Return accumulated session findings + visual evidence
return {
  findings: smart.findings(),
  scaffold: comparison.scaffold,
  gaps: { acme: comparison.siteA.gaps, rival: comparison.siteB.gaps },
  visualEvidence: { acme: acmeVisuals.sectionCount + ' sections captured' },
};

Examples

Navigate, extract, and analyze

// Connect to active tab, extract structured page evidence
const page = await smart.extractPage();
const finding = smart.finding({
  claim: `Site uses ${page.capture.framework?.name || 'no detected framework'}`,
  evidence: [`Framework: ${JSON.stringify(page.capture.framework)}`],
  sourceUrl: page.meta?.canonical || 'active tab',
  confidence: 'high',
  method: 'extractPage framework detection',
});
return { page: page.meta, finding };

Mobile emulation + screenshot

// Emulate iPhone and capture
await bridge.send({ type: 'emulate_device', device: 'iPhone 14' }, 10000);
await smart.navigate('https://example.com');
await smart.waitForIdle();
const screenshot = await bridge.send({ type: 'take_screenshot' }, 10000);
return screenshot;

Record and compile automation

// Record a browser session, then compile to reusable skill
await bridge.send({ type: 'start_recording' }, 10000);
await smart.navigate('https://example.com');
await smart.click('button.submit');
await smart.type('#email', 'test@example.com');
const session = await bridge.send({ type: 'stop_recording' }, 10000);
return compileRecording(session.session, 'signup-flow');

Intercept and mock network

// Block analytics, mock API response
await bridge.send({
  type: 'browser_intercept',
  pattern: '*analytics*',
  action: 'block'
}, 10000);
await bridge.send({
  type: 'browser_intercept',
  pattern: '*/api/user',
  action: 'mock',
  body: JSON.stringify({ name: 'Test User' }),
  statusCode: 200
}, 10000);
await smart.navigate('https://example.com');
return await smart.snapshot();

Session Recording

Record browser sessions as structured data, then compile them into reusable automation skills. 12 interaction tools are automatically intercepted during recording (click, type, navigate, scroll, etc.), capturing args, result, timing, and page URL.

// In code mode: record, interact, compile
await bridge.send({ type: 'start_recording' }, 10000);
// ... interact with the page ...
const session = await bridge.send({ type: 'stop_recording' }, 10000);
const skill = compileRecording(session.session, 'my-automation');
return skill;

In full mode, recording is available as 4 individual tools: start_recording, stop_recording, get_recording_status, and compile_recording.

Auto-Settling

Mutative tools (browser_click, browser_type, browser_navigate, browser_select_option) use actionability checks:

1. Pre-flight: Polls element visibility, stability, and enabled state before acting
2. Action: Dispatches the CDP command
3. Post-settle: Waits for DOM mutations to quiesce with progressive backoff [0, 20, 100, 100, 500]ms

This means the AI doesn't need to manually add sleep() or waitFor() calls between actions — the tools handle SPA rendering delays automatically.

Framework Detection

Detects 64 technologies across 4 tiers using globals, DOM markers, meta tags, HTTP headers, and script URLs:

Multi-framework detection returns a primary framework (meta-framework takes priority) plus a subFrameworks array for the full stack.

Tools Reference

The table below is a hand-written guide to the commonly used tools, not the complete set. For the full, current surface — all 145 of them, read from the server itself — run npx crawlio-browser tools --full.

Requirements

• Node.js >= 18
• Chrome (or Chromium) with the Crawlio for Chrome extension installed
• Crawlio.app (optional) — for site crawling and enrichment

Build from Source

git clone https://github.com/Crawlio-app/crawlio-browser.git
cd crawlio-browser
npm install
npm run build          # selectors → semantic-grounding → MCP server → extension

Build output lands in dist/mcp-server/ (ESM bundle) and dist/extension/ (the unpacked MV3 extension).

To develop against the extension, build the dev variant and load it unpacked:

npm run build:dev      # → dist/extension-dev, with __DEV__ logging enabled

Then open chrome://extensions, enable Developer mode, choose Load unpacked, and select dist/extension-dev. Reload the extension there after each rebuild.

npm test               # vitest
npm run typecheck      # server + extension TypeScript projects

Resources

• Documentation
• API Reference
• Product Page
• Chrome Extension
• npm Package
• Source
• Changelog | Releases

Contributing

Contributions are welcome — see CONTRIBUTING.md for the build and development loop, and CODE_OF_CONDUCT.md for community expectations. Please report security vulnerabilities privately via SECURITY.md rather than in a public issue.

License

Apache-2.0 — see LICENSE.

The selector kernel in packages/selectors is MIT-licensed: it contains code ported from Selector Forge (MIT) and Chromium DevTools (BSD-3-Clause). See packages/selectors/LICENSE and THIRD_PARTY_NOTICES.md for full attributions.