custom-browser-mcp
MCP server that extracts accessibility tree, design tokens, screenshots, Claude DSL, and Figma JSON from any URL using a persistent Chromium browser with zero LLM cost.
README
eclectique-browser-mcp
Eclectique Browser MCP — turns any live URL into a multi-output design bundle: a11y tree, design tokens, screenshot, Claude-consumable DSL, Figma REST JSON. Zero LLM API cost. Persistent Chromium context.
Tools (3)
analyze_page
Multi-output bundle. Pick any subset of {a11y, tokens, screenshot}.
// input
{
"url": "https://example.com",
"outputs": ["a11y", "tokens", "screenshot"],
"viewport": { "width": 1440, "height": 900 },
"full_page_screenshot": true,
"wait_until": "networkidle",
"timeout_ms": 30000
}
Output: Bundle v1.0.0 (schema). Versioned snapshot_id (sha256 url+ts).
| Field | Type | Notes |
|---|---|---|
a11y.yaml |
string | Playwright ariaSnapshot() YAML |
a11y.node_count / approx_tokens |
int | metrics |
tokens.colors_top |
array | top 16 frequencies with confidence: high|medium|low |
tokens.fonts |
array | family + weights + sample_count |
tokens.spacing_scale / type_scale_px |
number[] | distinct values |
screenshot.path / bytes |
string / int | PNG full-page on disk |
warnings |
string[] | per-extractor failures (non-fatal) |
to_claude
cbm/htmltoclaude/v0 — compact YAML DSL. Token-economical for Claude ingestion.
// input
{ "url": "https://example.com", "format": "yaml" }
Output: YAML with hoisted token refs (c0..cN colors, f0..fN fonts), 4-tuple boxes, node types FRAME|TEXT|IMAGE|SVG|GROUP|INPUT, conservative pruning of empty GROUPs, shadowRoot recursion, warnings for iframe_cross_origin / bg_gradient_unmapped / shadow_dom_skipped.
Sample: examples/awwwards-sotd.htmltoclaude.yaml (180 nodes, ~19.5K tokens).
to_figma
cbm/htmltofigma/v0 — Figma REST node JSON. Custom adapter (no proprietary API).
// input
{ "url": "https://example.com" }
Output: { document: { children: [{ type: "CANVAS", children: [FigmaNode...] }] }, warnings, metrics }. Mapping table:
| ClaudeNode | FigmaNode | Coverage v0 |
|---|---|---|
| FRAME | FRAME | fill SOLID + corner radius + border + clip |
| TEXT | TEXT | characters + style (fontFamily/Size/Weight) |
| IMAGE | RECTANGLE | fill IMAGE (imageRef = src, scaleMode) |
| SVG | VECTOR | shell only — path data deferred Phase 3 |
| GROUP | GROUP | container |
| INPUT | FRAME | placeholder |
Sample: examples/awwwards-sotd.htmltofigma.json (180 nodes, ~174 KB JSON).
Deferred Phase 3+: SVG paths, gradients (GRADIENT_LINEAR), box shadows (DROP_SHADOW), auto-layout, COMPONENT/INSTANCE detection.
Install
bun install
bunx playwright install chromium-headless-shell
Run
bun run serve # stdio MCP server
bun run smoke [url] # analyze_page smoke
bun run smoke:claude [url] # to_claude smoke (writes out/htmltoclaude-{N}n.yaml)
bun run smoke:figma [url] # to_figma smoke (writes out/htmltofigma-{N}n.json)
bun run consolidate [url] # sequential 3-tool + assert htmltoclaude↔figma node parity
bun run typecheck # tsc --noEmit
Wire into Claude Code (.mcp.json)
{
"mcpServers": {
"eclectique-browser": {
"command": "bun",
"args": ["run", "serve"],
"cwd": "/absolute/path/to/eclectique-browser-mcp"
}
}
}
Restart Claude Code. Verify with /mcp — should list eclectique-browser with 3 tools.
Stack
- Bun 1.3 — runtime + package manager
- TypeScript 5
- @modelcontextprotocol/sdk 1.29 — stdio transport
- Playwright 1.60 — Chromium headless persistent context
- Zod 4 — input/output schema validation
- yaml 2.9 — YAML serialization (
to_claude)
Architecture
src/
├── lib/
│ └── browser.ts ← shared Chromium persistent context (singleton, viewport-aware)
├── extractors/
│ ├── a11y.ts ← page.locator(':root').ariaSnapshot() YAML
│ ├── tokens.ts ← getComputedStyle walk → colors/fonts/spacing/type
│ ├── screenshot.ts ← page.screenshot({ fullPage })
│ ├── htmltoclaude.ts ← in-browser walker + token hoist + conservative pruning
│ └── figma.ts ← ClaudeBundle → FigmaDocument pure transform
├── tools/
│ ├── analyze_page.ts
│ ├── to_claude.ts
│ └── to_figma.ts
├── schemas/
│ └── output.ts ← Zod Bundle v1.0.0
└── server.ts ← MCP stdio + 3 tool registrations
Single Chromium persistent context shared across all 3 tools (factory in src/lib/browser.ts). Profile: out/.chromium-profile/ (override via CBM_USER_DATA_DIR env). Warm context ≈ 25% speedup vs cold launch.
Browser mode (3 options)
Set via CBM_BROWSER_MODE env var. Defaults to chromium.
| Mode | What | Pre-req | Use when |
|---|---|---|---|
chromium (default) |
Playwright-managed Chromium download (~92 MB once) | none | First install, CI, reproducible runs |
chrome |
Use Google Chrome stable installed on machine | Chrome installed | Skip 92 MB download, auto-update via Chrome |
cdp |
Connect to running Chrome via DevTools Protocol | Launch Chrome with --remote-debugging-port=9222 |
Share user cookies/sessions (logged-in IG/Gmail), anti-bot maximal |
# Mode chromium (default — managed download)
bun run smoke https://example.com
# Mode chrome (use system Chrome)
CBM_BROWSER_MODE=chrome bun run smoke https://example.com
# Mode cdp (connect to running Chrome with debug port)
# Step 1: launch Chrome with debug port
google-chrome --remote-debugging-port=9222
# Step 2: run with CDP mode
CBM_BROWSER_MODE=cdp CBM_CDP_URL=http://localhost:9222 bun run smoke https://example.com
CDP mode shares ALL browser state (cookies, localStorage, logged-in accounts) — useful for scraping behind login walls, but privacy risk : MCP sees all your tabs and sessions.
Cost & performance
Measured smoke baseline against Awwwards SOTD (1440×900, networkidle), Mac arm64:
| Tool | Cold | Warm | Output | LLM cost |
|---|---|---|---|---|
analyze_page (all outputs) |
~6.5s | ~2.3s | 159 a11y nodes + 5 colors + 453 KB PNG | $0 |
to_claude |
~1.9s | ~1.3s | 180 nodes / 19.5K YAML tokens | $0 |
to_figma |
~1.9s | ~1.3s | 180 nodes / 174 KB JSON / 43K tokens | $0 |
| Consolidate (all 3) | — | ~4.9s | shared context warm | $0 |
Phase 1 + 2 are 100% local Chromium. Optional LLM augmentation (vision tagging, code emit) gated Phase 3+.
License
Status
v0.4.2 — Phase 1 (analyze_page) + Phase 1b (to_claude) + Phase 2 v0 (to_figma) + cookie consent lifecycle shipped.
Deferred coverage (Phase 3+): SVG path traversal, gradient parser (GRADIENT_LINEAR), box-shadow effects (DROP_SHADOW), auto-layout heuristic detection (flex/grid), COMPONENT/INSTANCE detection, multi-tab, vision-LLM augmentation.
Issues / PRs welcome.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.