🕷️ Arachne MCP Server

🌐 English · 🇧🇷 Português

15 MCP tools for web scraping, browser automation, computer vision, audio transcription, RAG, and Visual Regression Testing (VRT) — the Oito Olhos (Spider's Eight Eyes) — all through the Arachne API.

Backed by Arachne. Works with Claude Desktop, Cursor, Codex CLI, Hermes, and any MCP client.

✨ Tools

🔍 Web & Data

👁️ Vision & Audio

🕷️ Oito Olhos — Visual Regression Testing (VRT)

🧭 Utilities

VRT with local LLM: Oito Olhos uses pixel diff + DOM pairing + CLIP semantic triage + local VLM (Qwen2.5-VL via Douglas, Samuel mirror fallback) — classifying changes as regression | expected | noise, with a fix suggestion (CORRECAO:) when it detects a bug. No third-party API key for the verdict.

🚀 Quick Start

1. Get an API key

Create one at arachne.seu.pet/dev (Free plan: 500 req/month).

2. Configure in Claude Desktop

{
  "mcpServers": {
    "arachne": {
      "command": "python3",
      "args": ["-m", "arachne_mcp"],
      "env": {
        "ARACHNE_API_KEY": "your_key_here",
        "ARACHNE_BASE_URL": "https://arachne.seu.pet"
      }
    }
  }
}

3. Or run directly

export ARACHNE_API_KEY="your_key"
python3 -m arachne_mcp

Requires httpx: pip install httpx

📦 How it works

The MCP server is an HTTP client that calls the public Arachne API. Zero local infrastructure — runs anywhere.

Your AI agent → MCP stdio → arachne_mcp.py → HTTP → Arachne API → result

🕷️ Examples — Oito Olhos (VRT)

# 1. Create a baseline for a page (first run)
arachne_visual_snapshot(url="https://mysite.com/", name="home")

# 2. After a deploy: compare and get a verdict
arachne_visual_diff(url="https://mysite.com/", name="home")
# → { diff_ratio: 0.023, classification: "regression",
#     description: "CTA button changed color. CLASSIFICACAO: regression
#     CORRECAO: check .cta CSS — color should be var(--primary)" }

# 3. Expected change? Approve as the new baseline
arachne_visual_approve(name="home")

# 4. Baseline-free audit (deterministic gates)
arachne_visual_gates(url="https://mysite.com/")
# → verdict: CLEAN | DEFECTS with machine-parsable fix list

# 5. Full suite → HTML report
arachne_visual_report(tests=[{"url": "https://mysite.com/", "name": "home"},
                             {"url": "https://mysite.com/pricing", "name": "pricing"}])

Tips:

• Authenticated pages: use auth: true — the engine logs in and injects the token.
• Dynamic areas (stats, clock): use mask: [{"selector": ".stats"}].
• Color/theme differences that aren't bugs: CLIP semantic triage classifies them as expected without invoking the 7B VLM.

📊 Plans

🏗️ Stack

• Backend: FastAPI + Crawl4AI + Whisper + Tesseract + PostgreSQL
• VRT: Playwright + ffmpeg (video) + CLIP ViT-B-32 + Qwen2.5-VL (local VLM)
• Engines: Trafilatura → Crawl4AI SDK → Sidecar Docker → Camoufox
• MCP Transport: stdio (compatible with Claude Desktop, Cursor, Codex, Hermes)

🔗 Links

• Arachne Platform
• Developer Portal
• Python SDK (PyPI)
• GitHub

🕷️ Built with the Arachne engine. Open source MCP server.