📑 Table of contents

• What is an MCP?
• The whole pitch in 4 lines
• One command = a complete audit
• Why this exists
• Install in 60 seconds — one-liner · Docker · manual
• 50+ checks every audit
• Feature comparison
• What every audit produces
• Your first audit
• How it works · Architecture
• Configuration
• 37 MCP tools
• Documentation
• Roadmap
• License & credits

🤔 Don't know what an MCP is? Read this 30-second explainer

Model Context Protocol (MCP) is the open standard that lets AI assistants like Claude, Cursor, or Codex call external tools. Think of it as "USB for AI assistants" — you plug a tool in, the AI can use it. librecrawl-technical-seo-audit-mcp is one of those tools. Once installed, you just ask your AI assistant to audit a site, and it does. No GUI. No dashboard. No exports.

New to all this?

• Don't have Claude Code yet? → Install Claude Code (free for individuals).
• Prefer Cursor? → Get Cursor.
• Already have one of those? → Skip to Install in 60s.

🪄 The whole pitch in 4 lines

You:    Audit https://acme.com — full site, no caps, give me the zip
Agent:  → librecrawl_start_chunked_audit · polls until done · saves zip locally
You:    Show me broken pages + broken external links + hreflang errors
Agent:  → reads CSVs, prints filtered tables. Server already forgot the audit.

That's the product. Your AI assistant runs a full technical SEO audit for you. You get a branded PDF + 7 CSVs covering 50+ technical checks, ready to hand a client. The server wipes everything the moment you download.

✅ One command = a complete audit (v2.2.0)

No flags to remember, no caps to set. Anyone who runs:

librecrawl_start_chunked_audit(url="https://yoursite.com")

gets a full end-to-end audit by default:

• Every page — entire sitemap crawled, including orphan pages not linked from the homepage
• Every page's text — readability, AI-tells, boilerplate analyzed word-by-word
• Every outbound link — every domain (yours, third-party, social, CDN) HTTP-validated for broken links
• No silent dropping — per-page checks + link validation cover 100% of pages; the report states exactly what was covered
• Heavy sites safe — 4–5 MB pages crawl fine; Screaming-Frog-grade politeness never overloads the origin
• Zero history — the server forgets the entire audit the moment you download the zip; re-scan anytime, nothing persists

Verified on real production sites: 1,942-page and 709-page WordPress sites, full coverage, origins healthy throughout.

Note: runs one audit at a time (single-tenant backend). For team use, queue audits sequentially — concurrent multi-audit routing is on the roadmap.

🔥 Why this exists

There are great desktop SEO crawlers (you know the ones). There are great cloud SEO suites. There was no AI-native crawler. librecrawl-technical-seo-audit-mcp fills that gap with five things no comparable open-source MCP server does:

⚡ It runs inside your AI assistant

37 MCP tools your agent calls directly. No GUI app to babysit, no SaaS dashboard to log into, no CSV exports to upload to ChatGPT. You just ask.

🚀 Chunked-progressive crawler that never times out

Most SEO MCP servers (SiteAudit MCP, AgentAEO, SE Ranking MCP) run synchronously and disconnect on sites over a few hundred pages. librecrawl-technical-seo-audit-mcp runs the crawl in a background worker thread, persists progress to SQLite WAL, and returns a session_id in under 2 seconds. Your agent polls a tiny status tool until done. 10,000-page enterprise sites work the same as 50-page blogs. Survives PM2 / MCP-client restarts mid-crawl.

🛡️ Catches WAF challenges other crawlers silently misreport

Cloudflare, Akamai, DataDome, Imperva, and PerimeterX challenge pages are served as 200 OK but contain a JavaScript challenge instead of your content. Most crawlers report these as "page OK, all good". librecrawl-technical-seo-audit-mcp fingerprints the challenge in the response body and flags bot_block_challenge_detected. You see what's actually broken.

🤖 An AIMD controller tunes crawl delay live

Additive-Increase / Multiplicative-Decrease — the same algorithm TCP congestion control uses. Error rate > 10% → halve chunk, double delay. p95 latency > 1.5× target → 1.5× delay. Clean signals → additive decrease. Polite by construction. No rate-limit blow-ups. No manual tuning. Respects robots.txt Crawl-Delay floor.

🧹 Ephemeral by design — the agency-safe default

Once you download the zip, the server deletes the session row, every artifact file on disk, AND the upstream LibreCrawl crawl record. Per-audit server footprint after cleanup: 0 bytes, 0 rows. Auditing 50 client sites? Zero data persists where another operator could see it.

📄 Branded PDF reports ready to hand a client

WeasyPrint, A4, page numbers, footer on every page. Open in any PDF viewer. No SaaS watermark. Hand it to a client as your work.

⚡ Install in 60 seconds

curl -fsSL https://raw.githubusercontent.com/adityaarsharma/librecrawl-technical-seo-audit-mcp/main/install.sh | bash

The installer asks 3 questions (target client, optional Google PageSpeed API key, optional GSC integration) and writes a ready-to-use MCP entry into your Claude / Cursor / Codex / Windsurf config. Done.

Three ways to run: the one-liner above (guided installer) · Docker (full stack, one command — best for self-hosting/VPS) · manual (Python venv + your own backend). New here? Start with the one-liner.

🐳 Run the full stack with Docker (recommended for self-hosting)

Brings up both the LibreCrawl engine and the MCP server, wired together and health-gated, with one command. Requires Docker + Docker Compose.

git clone https://github.com/adityaarsharma/librecrawl-technical-seo-audit-mcp.git
cd librecrawl-technical-seo-audit-mcp
docker compose up --build          # MCP at http://127.0.0.1:5081/mcp

The mcp service waits for LibreCrawl to report healthy before it starts. First build takes a few minutes (LibreCrawl pulls Chromium for JS rendering). Audit zips land in ./reports. Full knobs in docs/CONFIGURATION.md.

Add to your client config (Claude Desktop / Code — HTTP transport):

{
  "mcpServers": {
    "librecrawl": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://127.0.0.1:5081/mcp"]
    }
  }
}

🔧 Manual install

Python 3.10+, plus a LibreCrawl backend (the Docker path above is the easiest way to get one).

git clone https://github.com/adityaarsharma/librecrawl-technical-seo-audit-mcp.git
cd librecrawl-technical-seo-audit-mcp
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# Point the MCP at your running LibreCrawl backend, then start it:
LIBRECRAWL_URL=http://127.0.0.1:5080 python server.py

Add to your client config (Claude Desktop example):

{
  "mcpServers": {
    "librecrawl": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://127.0.0.1:5081/mcp"]
    }
  }
}

Per-client setup (Claude Code, Cursor, Windsurf, Codex, Continue.dev) and every environment variable: docs/GETTING-STARTED.md · docs/CONFIGURATION.md.

🚀 50+ checks every audit

🔗 Every outbound URL HEAD/GET-validated into 17 status classes — ok · redirect · forbidden · not_found · timeout · dns_error · ssl_error · connection_refused · etc. Per-target: final URL after redirects, source pages, anchor text, response time, server header.

📈 GSC merge — pull Google Search Console data, call librecrawl_merge_gsc_data(crawl_id, gsc_data). URLs normalised before joining. Emits 4 extra CSVs: per-page-with-gsc · gsc-winners · gsc-losers (high impr + CTR <2%) · gsc-quick-wins (position 11–20 + impr ≥100).

🆚 Feature comparison to other on-site SEO crawlers

This is a factual feature comparison. Prices were checked at publication and may have changed — see each vendor's site for current pricing. Brand names belong to their respective owners.

Reading guide: if you currently use a paid on-site crawler and your workflow is "crawl → export CSVs → analyse", librecrawl-technical-seo-audit-mcp covers that flow inside your AI assistant for £0 with no page caps. If your workflow depends on JavaScript-rendered SPAs, that's on the roadmap but not shipped yet — use the desktop tool for now.

📊 What every audit produces

Single zip, 8 files:

📖 Your first audit

You:   Audit https://example.com — full site, no caps

Agent: → librecrawl_start_chunked_audit(url=..., total_max_pages=10000)
         returns session_id in <2s

       → polls librecrawl_audit_status every 25s
         status: crawling, pages_done: 47,  current_delay_ms: 250
         status: crawling, pages_done: 312, last chunk p95: 480ms, err_rate: 0%
         status: done,     pages_done: 534, artifacts_ready: true

       → librecrawl_audit_zip(session_id, auto_cleanup=True)
         returns base64 zip (8 files, 320 KB)
         SAVES LOCALLY as example.com-1780572742.zip
         Server wiped: session_rows=4, files=8, upstream_crawl=1

You:   Show me broken pages + broken external links

Agent: → unzips, reads per-page.csv (filters status_4xx OR status_5xx)
       → reads external-links.csv (filters not_found · forbidden · 5xx · timeout)
       → prints both tables

Local zip is the only copy. Server is back to zero state.

🛣️ Roadmap

Not planned: keyword research, backlink analysis, SERP tracking. Those are different problems with different MCP servers (DataForSEO, etc.). This tool is laser-focused on technical on-site SEO crawling.

Open an issue to bump priorities or request a check.

🔎 How it works

One audit, start to finished zip — every step is an MCP tool your agent calls:

sequenceDiagram
    actor You
    participant Agent as AI assistant
    participant MCP as librecrawl-mcp
    participant LC as LibreCrawl engine
    You->>Agent: "Audit https://acme.com — full site, give me the zip"
    Agent->>MCP: librecrawl_start_chunked_audit(url)
    MCP->>LC: start crawl on a background worker
    MCP-->>Agent: session_id (returned in < 2s)
    loop every ~25s until done
        Agent->>MCP: librecrawl_audit_status(session_id)
        MCP-->>Agent: crawling · pages_done · p95 latency · err_rate
    end
    MCP-->>Agent: status: done · artifacts_ready
    Agent->>MCP: librecrawl_audit_zip(session_id, auto_cleanup=True)
    MCP-->>Agent: base64 zip — branded PDF + 7 CSVs
    Agent-->>You: saved locally · server wiped to 0 bytes, 0 rows

🏗️ Architecture

Two processes: a thin MCP wrapper your agent talks to, and the LibreCrawl engine that does the crawling.

flowchart TD
    A["🖥️ MCP client<br/>Claude Code · Desktop · Cursor · Codex · Windsurf"]
    A -->|streamable HTTP or stdio| B
    subgraph MCP["🕷️ librecrawl-technical-seo-audit-mcp — server.py (FastMCP · 37 tools)"]
        direction TB
        B["runner.py<br/>background worker thread + AIMD controller"]
        C["state.py<br/>SQLite WAL — session state"]
        D["external_links · schema · content_audit<br/>extended_checks · sitemap_fill · pdf_report"]
        E["libreclient.py<br/>typed wrapper to the upstream API"]
        B --> C
        B --> D
        B --> E
    end
    E -->|HTTP :5080| F["🐍 LibreCrawl Flask backend<br/>single-tenant · crawls + extracts SEO data"]

Full component walk-through: docs/ARCHITECTURE.md.

⚙️ Configuration

All environment variables are optional — the defaults just work. Set them via your shell, docker compose, or your MCP client config.

📖 Full reference, per-client config, and transport details: docs/CONFIGURATION.md.

🛠️ 37 MCP tools

📚 Documentation

Deeper guides live in docs/:

📜 License & trademarks

Code: MIT. Use it on client work, agency work, internal tools, anything. No attribution required (but appreciated). See LICENSE.

Trademarks. All third-party product names mentioned in this README (including any names referenced in the comparison table) are property of their respective owners. This project is not affiliated with, endorsed by, or sponsored by any third-party tool vendor. Comparisons are based on publicly available information at the time of writing and exist for the purpose of informing readers evaluating different categories of SEO tooling.

🙏 Credits

• LibreCrawl — the upstream open-source crawler this MCP server wraps. MIT. Please go star them — this project would not exist without that work.
• Anthropic Model Context Protocol — the protocol this server speaks
• WeasyPrint — Markdown → HTML → PDF rendering
• FastMCP — the Python MCP server framework