Charlotte

The Web, Readable.

Your AI agent spends 60,000 tokens just to look at a web page. Charlotte does it in 336.

Charlotte is an MCP server that gives AI agents structured, token-efficient access to the web. Instead of dumping the full accessibility tree on every call, Charlotte returns only what the agent needs: a compact page summary on arrival, targeted queries for specific elements, and full detail only when explicitly requested. The result is 25-182x less data per page compared to Playwright MCP, saving thousands of dollars across production workloads.

Why Charlotte?

Most browser MCP servers dump the entire accessibility tree on every call — a flat text blob that can exceed a million characters on content-heavy pages. Agents pay for all of it whether they need it or not.

Charlotte decomposes each page into a typed, structured representation — landmarks, headings, interactive elements, forms, content summaries — and lets agents control how much they receive with three detail levels. When an agent navigates to a new page, it gets a compact orientation (336 characters for Hacker News) instead of the full element dump (61,000+ characters). When it needs specifics, it asks for them.

Benchmarks

Charlotte v0.6.2 vs Playwright MCP, measured by characters returned per tool call on real websites:

Navigation (first contact with a page):

Site	Charlotte navigate	Playwright browser_navigate
example.com	612	817
Wikipedia (AI article)	7,667	1,040,636
Hacker News	336	61,230
GitHub repo	3,185	80,297

Charlotte's navigate returns minimal detail by default — landmarks, headings, and interactive element counts grouped by page region. Enough to orient, not enough to overwhelm. On Wikipedia, that's 135x smaller than Playwright's response.

Tool definition overhead (invisible cost per API call):

Profile	Tools	Def. tokens/call	Savings vs full
full	43	~7,600	—
browse (default)	23	~3,900	~49%
core	7	1,677	~78%

Tool definitions are sent on every API round-trip. With the default browse profile, Charlotte carries ~49% less definition overhead than loading all tools. Over a 20-call browsing session, that's ~40% fewer total tokens. See the profile benchmark report for full results.

The workflow difference: Playwright agents receive 61K+ characters every time they look at Hacker News, whether they're reading headlines or looking for a login button. Charlotte agents get 336 characters on arrival, call find({ type: "link", text: "login" }) to get exactly what they need, and never pay for the rest.

How It Works

Charlotte maintains a persistent headless Chromium session and acts as a translation layer between the visual web and the agent's text-native reasoning. Every page is decomposed into a structured representation:

┌─────────────┐     MCP Protocol     ┌──────────────────┐
│   AI Agent  │<────────────────────>│    Charlotte     │
└─────────────┘                      │                  │
                                     │  ┌────────────┐  │
                                     │  │  Renderer  │  │
                                     │  │  Pipeline  │  │
                                     │  └─────┬──────┘  │
                                     │        │         │
                                     │  ┌─────▼──────┐  │
                                     │  │  Headless  │  │
                                     │  │  Chromium  │  │
                                     │  └────────────┘  │
                                     └──────────────────┘

Agents receive landmarks, headings, interactive elements with typed metadata, bounding boxes, form structures, and content summaries — all derived from what the browser already knows about every page.

Features

Navigation — navigate, back, forward, reload

Observation — observe (3 detail levels, structural tree view), find (spatial + semantic search, CSS selector mode), screenshot (with persistent artifact management), screenshots, screenshot_get, screenshot_delete, diff (structural comparison against snapshots)

Interaction (iframe-aware) — click, click_at (coordinate-based), type (with slow typing support), select, toggle, submit, scroll, hover, drag, key (single/sequence with element targeting), wait_for (async condition polling), upload (file input), fill_form (batch form fill), dialog (accept/dismiss JS dialogs)

Monitoring — console (all severity levels, filtering, timestamps), requests (full HTTP history, method/status/resource type filtering)

Session Management — tabs, tab_open, tab_switch, tab_close, viewport (device presets), network (throttling, URL blocking), set_cookies, get_cookies, clear_cookies, set_headers, configure

Development Mode — dev_serve (static server + file watching with auto-reload), dev_inject (CSS/JS injection), dev_audit (a11y, performance, SEO, contrast, broken links)

Utilities — evaluate (arbitrary JS execution in page context)

Tool Profiles

Charlotte ships 43 tools (42 registered + the charlotte_tools meta-tool), but most workflows only need a subset. Startup profiles control which tools load into the agent's context, reducing definition overhead by up to 78%.

charlotte --profile browse    # 23 tools (default) — navigate, observe, interact, tabs
charlotte --profile core      # 7 tools — navigate, observe, find, click, type, submit
charlotte --profile full      # 43 tools — everything
charlotte --profile interact  # 31 tools — full interaction + dialog + evaluate
charlotte --profile develop   # 34 tools — interact + dev_serve, dev_inject, dev_audit
charlotte --profile audit     # 14 tools — navigation + observation + dev_audit + viewport

Agents can activate more tools mid-session without restarting:

charlotte_tools enable dev_mode    → activates dev_serve, dev_audit, dev_inject
charlotte_tools disable dev_mode   → deactivates them
charlotte_tools list               → see what's loaded

Quick Start

Prerequisites

• Node.js >= 20
• npm

Installation

Charlotte is listed on the MCP Registry as io.github.TickTockBent/charlotte and published on npm as @ticktockbent/charlotte:

npm install -g @ticktockbent/charlotte

Docker images are available on Docker Hub and GitHub Container Registry:

# Alpine (default, smaller)
docker pull ticktockbent/charlotte:alpine

# Debian (if you need glibc compatibility)
docker pull ticktockbent/charlotte:debian

# Or from GHCR
docker pull ghcr.io/ticktockbent/charlotte:latest

Or install from source:

git clone https://github.com/ticktockbent/charlotte.git
cd charlotte
npm install
npm run build

Run

Charlotte communicates over stdio using the MCP protocol:

# If installed globally (default browse profile)
charlotte

# With a specific profile
charlotte --profile core

# If installed from source
npm start

MCP Client Configuration

Claude Code

Create .mcp.json in your project root:

{
  "mcpServers": {
    "charlotte": {
      "type": "stdio",
      "command": "npx",
      "args": ["@ticktockbent/charlotte"],
      "env": {}
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "charlotte": {
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "charlotte": {
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "charlotte": {
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

VS Code (Copilot)

Add to .vscode/mcp.json:

{
  "servers": {
    "charlotte": {
      "type": "stdio",
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

Cline

Add to Cline MCP settings (via the Cline sidebar > MCP Servers > Configure):

{
  "mcpServers": {
    "charlotte": {
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

Amp

Add to ~/.amp/settings.json:

{
  "mcpServers": {
    "charlotte": {
      "command": "npx",
      "args": ["@ticktockbent/charlotte"]
    }
  }
}

See docs/mcp-setup.md for the full setup guide, including development mode, generic MCP clients, verification steps, and troubleshooting.

Usage Examples

Once connected, an agent can use Charlotte's tools:

Browse a website

navigate({ url: "https://example.com" })
// → 612 chars: landmarks, headings, interactive element counts

find({ type: "link", text: "More information" })
// → just the matching element with its ID

click({ element_id: "lnk-a3f1" })

Fill out a form

navigate({ url: "https://httpbin.org/forms/post" })
find({ type: "text_input" })
type({ element_id: "inp-c7e2", text: "hello@example.com" })
select({ element_id: "sel-e8a3", value: "option-2" })
submit({ form_id: "frm-b1d4" })

Local development feedback loop

dev_serve({ path: "./my-site", watch: true })
observe({ detail: "full" })
dev_audit({ checks: ["a11y", "contrast"] })
dev_inject({ css: "body { font-size: 18px; }" })

Page Representation

Charlotte returns structured representations with three detail levels that let agents control how much context they consume:

Minimal (default for navigate)

Landmarks, headings, and interactive element counts grouped by page region. Designed for orientation — "what's on this page?" — without listing every element.

{
  "url": "https://news.ycombinator.com",
  "title": "Hacker News",
  "viewport": { "width": 1280, "height": 720 },
  "structure": {
    "headings": [{ "level": 1, "text": "Hacker News", "id": "h-a1b2" }]
  },
  "interactive_summary": {
    "total": 93,
    "by_landmark": {
      "(page root)": { "link": 91, "text_input": 1, "button": 1 }
    }
  }
}

Summary (default for observe)

Full interactive element list with typed metadata, form structures, and content summaries.

{
  "url": "https://example.com/dashboard",
  "title": "Dashboard",
  "viewport": { "width": 1280, "height": 720 },
  "structure": {
    "landmarks": [
      { "id": "rgn-b2c1", "role": "banner", "label": "Site header", "bounds": { "x": 0, "y": 0, "w": 1280, "h": 64 } },
      { "id": "rgn-d4e5", "role": "main", "label": "Content", "bounds": { "x": 240, "y": 64, "w": 1040, "h": 656 } }
    ],
    "headings": [{ "level": 1, "text": "Dashboard", "id": "h-1a2b" }],
    "content_summary": "main: 2 headings, 5 links, 1 form"
  },
  "interactive": [
    {
      "id": "btn-a3f1",
      "type": "button",
      "label": "Create Project",
      "bounds": { "x": 960, "y": 80, "w": 160, "h": 40 },
      "state": {}
    }
  ],
  "forms": []
}

Full

Everything in summary, plus all visible text content on the page.

Detail Levels

Level	Tokens	Use case
minimal	~50-200	Orientation after navigation. What regions exist? How many interactive elements?
summary	~500-5000	Working with the page. Full element list, form structures, content summaries.
full	variable	Reading page content. All visible text included.

Navigation tools default to minimal. The observe tool defaults to summary. Both accept an optional detail parameter to override.

Element IDs

Element IDs are stable across minor DOM mutations. They're generated by hashing a composite key of element type, ARIA role, accessible name, and DOM path signature:

btn-a3f1  (button)    inp-c7e2  (text input)
lnk-d4b9  (link)      sel-e8a3  (select)
chk-f1a2  (checkbox)  frm-b1d4  (form)
rgn-e0d2  (landmark)  hdg-0f40  (heading)
dom-b2c3  (DOM element, from CSS selector queries)

IDs survive unrelated DOM changes and element reordering within the same container. When an agent navigates at minimal detail (no individual element IDs), it uses find to locate elements by text, type, or spatial proximity — the returned elements include IDs ready for interaction.

Development

# Run in watch mode
npm run dev

# Run all tests
npm test

# Run only unit tests
npm run test:unit

# Run only integration tests
npm run test:integration

# Type check
npx tsc --noEmit

Project Structure

src/
  browser/          # Puppeteer lifecycle, tab management, CDP sessions
  renderer/         # Accessibility tree extraction, layout, content, element IDs
  state/            # Snapshot store, structural differ
  tools/            # MCP tool definitions (navigation, observation, interaction, session, dev-mode)
  dev/              # Static server, file watcher, auditor
  types/            # TypeScript interfaces
  utils/            # Logger, hash, wait utilities
tests/
  unit/             # Fast tests with mocks
  integration/      # Full Puppeteer tests against fixture HTML
  fixtures/pages/   # Test HTML files

Architecture

The Renderer Pipeline is the core — it calls extractors in order and assembles a PageRepresentation:

1. Accessibility tree extraction (CDP Accessibility.getFullAXTree)
2. Layout extraction (CDP DOM.getBoxModel)
3. Landmark, heading, interactive element, and content extraction
4. Element ID generation (hash-based, stable across re-renders)

All tools go through renderActivePage() which handles snapshots, reload events, dialog detection, and response formatting.

Sandbox

Charlotte includes a test website in tests/sandbox/ that exercises all tools without touching the public internet. Serve it locally with:

dev_serve({ path: "tests/sandbox" })

Five pages cover navigation, forms, interactive elements, popups, delayed content, scroll containers, and more. See docs/sandbox.md for the full page reference and a tool-by-tool exercise checklist.

Known Issues

Shadow DOM — Open shadow DOM works transparently. Chromium's accessibility tree pierces open shadow boundaries, so web components (e.g., GitHub's <relative-time>, <tool-tip>) render their content into Charlotte's representation without special handling. Closed shadow roots are opaque to the accessibility tree and will not be captured.

Roadmap

Session & Configuration

Persistent Init Scripts — Add a --init-script CLI argument to inject JavaScript on every page load via page.evaluateOnNewDocument(). Charlotte's dev_inject currently applies CSS/JS once and does not persist across navigations.

Configuration File — Support a --config CLI argument to load settings from a JSON file, simplifying repeatable setups and CI/CD integration.

Full Device Emulation — Extend charlotte_viewport to accept named devices (e.g., "iPhone 15") and configure user agent, touch support, and device pixel ratio via CDP, not just viewport dimensions.

Feature Roadmap

Video Recording — Record interactions as video, capturing the full sequence of agent-driven navigation and manipulation for debugging, documentation, and review.

ARM64 Docker Images — Add linux/arm64 platform support to the Docker publish workflow for native performance on Apple Silicon Macs and ARM servers.

See docs/playwright-mcp-gap-analysis.md for the full gap analysis against Playwright MCP, including lower-priority items (vision tools, testing/verification, tracing, transport, security) and areas where Charlotte has advantages.

Full Specification

See docs/CHARLOTTE_SPEC.md for the complete specification including all tool parameters, the page representation format, element identity strategy, and architecture details.

License

MIT

Community

• Open a bug report for reproducible defects, regressions, or MCP-client-specific problems.
• Open a feature request for workflow improvements or new capabilities.
• Open a tool request if you want to propose a new tool, parameter surface, or profile placement.
• Browse open issues to find current work and discussion.
• Check the planned good first issue filter as maintainers tag starter-friendly tasks.

Contributing

See CONTRIBUTING.md for guidelines.

Part of a growing suite of literary-named MCP servers. See more at github.com/TickTockBent.