Leapfrog

Multi-session browser MCP for AI agents.
36 tools. 15 parallel sessions. Stealth. HUD. Self-improvement. Up to 10x fewer tokens.

npm i leapfrog-mcp  |  Works with Claude Code, Cursor, Windsurf

The Problem

Playwright MCP sends ~14,000 tokens for a content-heavy page like Hacker News. Most of that is noise. Your context window fills up. Your agent gets confused. You pay for it.

Leapfrog sends ~1,400 tokens. Same page. Same information. Up to 10x less noise.

┌─────────────────────────────────────────────────────┐
│  Playwright MCP                                     │
│  ████████████████████████████████████████  ~14,000   │
│                                                     │
│  Leapfrog                                           │
│  █████                                    ~1,400    │
└─────────────────────────────────────────────────────┘
          tokens per page (Hacker News, real test)

Savings range from 2-10x depending on page complexity. Content-heavy pages see the biggest wins. Dense forms see the smallest. The median across real-world sites is ~4-5x.

Quick Start

npx leapfrog-mcp --doctor          # verify everything works
npx leapfrog-mcp --stealth-audit   # test all 19 stealth patches
npx leapfrog-mcp --config          # print MCP config to paste

Add to ~/.mcp.json (Claude Code) or your editor's MCP config:

{
  "leapfrog": {
    "command": "npx",
    "args": ["-y", "leapfrog-mcp"],
    "env": {
      "LEAP_MAX_SESSIONS": "15",
      "LEAP_TILE": "true",
      "LEAP_HUD": "true",
      "LEAP_AUTO_CONSENT": "true"
    }
  }
}

Leapfrog uses playwright-core (15MB) instead of playwright (1.6GB) and does not bundle a browser. Either:

• Set LEAP_CHANNEL=chrome to use your installed Chrome/Chromium (recommended)
• Or run npx playwright-core install chromium to install the bundled Chromium binary
• Or set LEAP_CDP_ENDPOINT to connect to an already-running Chrome instance

Feature Matrix

Stealth

Leapfrog ships 19 anti-detection patches enabled by default (LEAP_STEALTH=true). Four modes:

• true (default) — all 19 patches active
• passive — removes automation signals only (webdriver, HeadlessChrome). Does NOT fake identity (WebGL, fonts, audio). Better for sites where trust matters more than evasion.
• auto — per-domain EXP3 bandit selects the optimal stealth configuration based on what's worked before
• false — no stealth patches

These cover the vectors that fingerprint services like CreepJS and fingerprint-pro actually check:

• Client Hints brands (strips HeadlessChrome)
• navigator.webdriver forced to undefined
• WebGL vendor/renderer (replaces SwiftShader with real GPU strings)
• Connection RTT (non-zero)
• Alert dismiss timing (human-speed delay)
• Window outer/inner height offset
• MIME type array population
• Platform inference from user agent
• chrome.app emulation
• iframe contentWindow protection
• Media codec spoofing (canPlayType)
• document.hasFocus() override
• Source URL comment stripping
• Custom UA + stealth coexistence (custom user agents no longer disable stealth context)
• CDP Runtime.enable detection (Error.prepareStackTrace filter)
• Permissions API spoofing (20+ permission types)
• AudioContext fingerprint noise (getChannelData/getFloatFrequencyData)
• WebRTC IP leak prevention (ICE candidate filtering)
• Font enumeration fingerprint spoofing

Per-session stealth control: pass stealth: false in session_create to disable for a specific session.

Humanization (Experimental)

Set LEAP_HUMANIZE=true to enable human-like browser interaction. This is opt-in and adds latency in exchange for more realistic behavior. Six modules:

• Mouse — Bezier curve paths with Fitts's Law timing and micro-tremor jitter
• Typing — Log-normal inter-key delays (200ms median), key dwell time, bigram-aware speed, rollover typing
• Scroll — Inertial simulation with ramp-up and momentum decay (touchpad/mouse-wheel physics)
• Pause — Inter-action "think" delays that simulate cognitive gaps between actions
• Fingerprint — Coherent browser fingerprint generation (platform, device memory, GPU, timezone)
• Utils — Shared math primitives (Box-Muller gaussian, distributions)

Page Classification

Every navigate and snapshot call automatically classifies the page type using weighted signal scoring (no LLM required). 18 types:

login · search-results · product · product-list · checkout · article · dashboard · form · error · challenge · landing · documentation · profile · media · feed · qa · ecommerce · unknown

Classification drives smarter snapshot extraction — login pages surface form fields, articles surface content, dashboards surface interactive elements.

Harness Intelligence

The harness tracks every action in a session and classifies outcomes:

• Action outcome classification — SUCCESS, SILENT_CLICK, NAVIGATION, WRONG_ELEMENT, BLOCKED, ERROR, PENDING
• Bot redirect detection — detects when a site redirects to a challenge or block page after an action
• Loop detection — warns when the agent is stuck clicking the same element, ping-ponging between URLs, or repeating actions
• Session memory — session_memory tool recalls actions after context window compression

Cookie Persistence

Persistent browser profiles now use context.cookies() + addCookies() instead of storageState(), which returns empty on persistent contexts. Auth state survives across sessions.

Adaptive Wait + Stealth Escalation

Navigate automatically retries with fallback strategies when pages fail to load:

1. Try load (fastest) — if empty, retry with networkidle (10s cap)
2. If networkidle times out (Amazon, ad-heavy sites), fall back to domcontentloaded
3. If blocked/challenged, escalate stealth: random delays → wait for JS challenge → rotate session with fresh fingerprint
4. Profile sessions (auth'd) never have their session destroyed — hard-capped at Level 2

Opt-out with autoRetry: false on navigate. Control max escalation with maxRetryLevel (0-5, default 3).

Record / Replay

Export a session's action history as a replayable recording, then replay it in new sessions:

• session_export — creates parameterized JSON or Playwright script from session history. @eN refs resolved to stable CSS selectors. Auto-detects emails, passwords, URLs as {{placeholders}}.
• session_replay — replays a recording with parameter overrides. Supports onError: 'stop' or 'skip'.

Turn one-off agent workflows into reusable automations.

Pagination Extraction

Extract data across multiple pages in a single tool call:

• Click-next — auto-detects "Next" buttons, pagination links, "Load more" buttons
• Infinite scroll — scrolls and waits for new content via DOM hash comparison
• URL pattern — increments ?page={page} or custom patterns

Replaces 3-4 tool calls per page. Cap: 50 pages, 100K total chars. Stops on: no next button, empty page, duplicate content, or bot detection.

Incremental Snapshots

The diff tool returns only what changed since the last snapshot — additions, removals, changes. Massive token savings for monitoring and polling workflows.

HUD Overlays (LEAP_HUD=true)

When running headed, Leapfrog overlays visual feedback on every session:

• Click ripple — expanding green circle at click coordinates (agent actions only)
• Zoom-to-target — browser zooms to 1.15x on the clicked element briefly so agents (and humans) can visually track what's happening in tiled windows
• Scroll-to-target — scrollIntoView before clicks so you can see what the agent is about to click

Minimal by design. No borders, no status bars, no cursor overlay — just the feedback that matters.

Multi-Terminal Tiling (LEAP_TILE=true)

Multiple Leapfrog instances share the screen via file-based coordination. Each instance tracks its own windows and a TilesCoordinator assigns global grid slots — no overlap, no manual arrangement. Set LEAP_TILE=true (or LEAP_TILE=master for the primary instance). Padding between tiles is configurable with LEAP_TILE_PADDING (default 8px).

Human Intervention

Leapfrog auto-detects situations that need a human — CAPTCHAs, login forms, OAuth redirects, Cloudflare challenges — and tries to self-resolve before pausing.