Cloudflare MCP Server for Static Sites

Turn your static website into an AI-accessible knowledge base. This Cloudflare Worker serves your content over MCP, so AI tools like Claude can search and retrieve it directly.

Cloudflare is well-suited for hosting remote MCP servers — its Workers platform handles the transport layer, and Durable Objects maintain persistent client sessions.

To see this approach in action, see my blog post Publishing Your Content to AI Assistants.

---

• Why This Matters
• How It Works
• Prerequisites
• Quick Start
• MCP Client Setup
• Threat Model
• Adapters
• Configuration
• Development
• Troubleshooting
• Examples
• Author

---

Why This Matters

AI assistants answer questions based on training data that may be outdated or incomplete. Web search and retrieval augmented generation (RAG) each put obstacles between the AI and your content. Web search depends on the AI choosing to search and hoping the right results come back. RAG can be tricky to deploy, and readers can't easily add it to their AI tools.

An MCP server makes your content a native capability of any AI tool that connects to it. The AI discovers and queries your content automatically, in a fast and token-friendly way.

You might use this to:

• Help users find answers in your documentation
• Give AI assistants access to your blog's content
• Let AI tools cite your articles with accurate, up-to-date information

How It Works

┌─────────────────────────────────────────────────────────────────────────┐
│                           Your Static Site                              │
│                    (Markdown files with frontmatter)                    │
└──────────────────────────────┬──────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                             Adapter                                     │
│         (Astro, Hugo, or Generic — runs at build time)                  │
│                                                                         │
│   Scans your content files, extracts metadata from frontmatter,         │
│   and generates a search-index.json file.                               │
└──────────────────────────────┬──────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                         Cloudflare R2                                   │
│                                                                         │
│   Stores the search index. Only your Worker can access it.              │
│   The Worker caches the index in memory for four hours.                 │
└──────────────────────────────┬──────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                       Cloudflare Worker                                 │
│                                                                         │
│   Implements the MCP server. Uses Fuse.js for fuzzy search.             │
│   Durable Objects maintain persistent sessions with MCP clients.        │
└──────────────────────────────┬──────────────────────────────────────────┘
                               │
                               ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                          MCP Clients                                    │
│           (Claude Desktop, Claude Code, Cursor, etc.)                   │
│                                                                         │
│   Tools available to the AI:                                            │
│   • search_<prefix> — Find content by keywords                          │
│   • get_article     — Retrieve a specific page by URL                   │
│   • get_index_info  — Get index statistics                              │
└─────────────────────────────────────────────────────────────────────────┘

Prerequisites

Requirement	What It's For
Cloudflare account	Hosts the Worker and R2 bucket. The free tier is sufficient.
Node.js 18+ or Bun	Runs the adapter that generates your search index.
Wrangler CLI	Deploys the Worker and manages R2. Installed via bun install.

Quick Start

You can follow these steps manually or point an AI coding tool (Claude Code, Cursor, etc.) at this repo and ask it to set things up. Either way, you'll need a Cloudflare account and these details about your site:

• Site name and domain (e.g., "My Blog" and "blog.example.com")
• Content directory path to your markdown files
• Tool prefix for MCP tool names (e.g., "myblog" → search_myblog)
• MCP endpoint domain (e.g., "mcp.example.com")

1. Clone and Install

git clone https://github.com/lennyzeltser/cloudflare-mcp-for-static-sites.git my-site-mcp
cd my-site-mcp
bun install

2. Configure

Edit wrangler.jsonc:

{
  "name": "my-site-mcp-server",
  "routes": [
    { "pattern": "mcp.example.com", "custom_domain": true }
  ],
  "r2_buckets": [
    { "binding": "SEARCH_BUCKET", "bucket_name": "my-site-mcp-data" }
  ]
}

3. Create R2 Bucket

bunx wrangler r2 bucket create my-site-mcp-data

4. Generate and Upload Index

Pick an adapter for your site (see Adapters):

bun adapters/generic/generate-index.js \
  --content-dir=../my-site/content \
  --site-name="My Site" \
  --site-domain="example.com" \
  --tool-prefix="mysite"

bunx wrangler r2 object put my-site-mcp-data/search-index.json \
  --file=./search-index.json \
  --content-type=application/json

5. Deploy

bun run deploy

Your MCP server is now running. Connect an MCP client to start searching.

CI/CD: The included GitHub Actions workflow (.github/workflows/deploy.yml) is set to manual trigger only. To deploy via GitHub Actions, go to Actions → Deploy → Run workflow. To enable auto-deploy on push, edit the workflow and add push: branches: [main] to the triggers.

---

MCP Client Setup

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "my-site": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.example.com/mcp"]
    }
  }
}

Claude Code

claude mcp add my-site --transport http https://mcp.example.com/mcp --scope user

Cursor

Add to your Cursor mcp.json:

{
  "mcpServers": {
    "my-site": {
      "url": "https://mcp.example.com/mcp"
    }
  }
}

Other Clients

Use the mcp-remote package to connect via the /mcp endpoint (streamable HTTP, recommended) or /sse endpoint (SSE transport, legacy).

Available Tools

Tool	Description
search_<prefix>	Search by keywords. Returns titles, URLs, dates, and summaries.
get_article	Retrieve full content by URL path (e.g., /about).
get_index_info	Get page count, generation date, and tool names.

---

Threat Model

This MCP server is designed for public content only. Consider these security characteristics before deploying:

What's Exposed

Exposure	Mechanism
All indexed content	get_article retrieves full text by URL path
Content enumeration	search_* with broad queries reveals page titles and summaries
Site metadata	/ endpoint and get_index_info reveal page count, domain, and tool names

Assumptions

• Your content is already public. The indexed pages come from a public website. This server makes them AI-searchable, not newly public.
• R2 is not the security boundary. While the R2 bucket is private, the Worker exposes its contents through MCP tools. Anyone with the endpoint URL can query all indexed content.
• No authentication. The MCP server accepts connections from any client. There's no API key, OAuth, or access control.

Not Designed For

• Private or internal documentation
• Content requiring authentication or authorization
• Partial access control (all-or-nothing visibility)

Recommendations

If you need access control, consider:

• Cloudflare Access for authentication at the Worker level
• A separate private deployment for internal content
• Excluding sensitive pages from the search index

---

Adapters

An adapter generates the search index from your content. It scans your files, extracts frontmatter metadata, and outputs search-index.json.

Each adapter handles the specifics of a particular static site generator.

Generic (Markdown)

Works with any site that uses markdown files with YAML frontmatter.

bun adapters/generic/generate-index.js \
  --content-dir=./content \
  --site-name="My Website" \
  --site-domain="example.com" \
  --tool-prefix="mysite" \
  --output=./search-index.json

See adapters/generic/README.md.

Astro

An Astro integration that generates the index at build time.

// astro.config.mjs
import { searchIndexIntegration } from './src/integrations/search-index.mjs';

export default defineConfig({
  integrations: [
    searchIndexIntegration({
      siteName: 'My Blog',
      siteDomain: 'blog.example.com',
      toolPrefix: 'myblog',
    }),
  ],
});

See adapters/astro/README.md.

Hugo

A Node.js script that handles both TOML and YAML frontmatter.

bun adapters/hugo/generate-index.js \
  --content-dir=./content \
  --site-name="My Hugo Site" \
  --site-domain="example.com"

See adapters/hugo/README.md.

Writing Your Own Adapter

If your static site generator isn't listed, you can write an adapter. It just needs to output JSON in the v3.0 format.

Your adapter should:

1. Find your content files (markdown, MDX, HTML, etc.)
2. Extract metadata from frontmatter (title, date, tags)
3. Extract body text for search
4. Map file paths to URLs
5. Write search-index.json

Here's a template:

import { writeFileSync } from 'fs';

const pages = [/* your content processing logic */];

const index = {
  version: "3.0",
  generated: new Date().toISOString(),
  site: {
    name: "My Site",
    domain: "example.com",
    description: "Brief description for the MCP tool",
    toolPrefix: "mysite",
  },
  pageCount: pages.length,
  pages: pages.map(page => ({
    url: page.url,           // Required: starts with /
    title: page.title,       // Required
    abstract: page.summary,  // Optional
    date: page.date,         // Optional: YYYY-MM-DD
    topics: page.tags,       // Optional: array
    body: page.content,      // Recommended for search quality
  })),
};

writeFileSync("search-index.json", JSON.stringify(index, null, 2));

Validate your index:

bun run validate ./search-index.json

Upload to R2:

bunx wrangler r2 object put my-site-mcp-data/search-index.json \
  --file=./search-index.json \
  --content-type=application/json

---

Configuration

wrangler.jsonc

Field	Description
name	Worker name in Cloudflare dashboard
routes[].pattern	Your custom domain
r2_buckets[].bucket_name	R2 bucket name

For testing, you can use a workers.dev subdomain instead of a custom domain:

"workers_dev": true,
// Comment out "routes"

Index Format

The search index follows the v3.0 schema:

{
  "version": "3.0",
  "generated": "2025-01-15T12:00:00.000Z",
  "site": {
    "name": "My Website",
    "domain": "example.com",
    "description": "A site about interesting topics",
    "toolPrefix": "mysite"
  },
  "pageCount": 42,
  "pages": [
    {
      "url": "/about",
      "title": "About Us",
      "abstract": "Learn about our team.",
      "date": "2025-01-01",
      "topics": ["about", "team"],
      "body": "Full page content..."
    }
  ]
}

Field	Required	Description
version	Yes	Schema version ("3.0")
generated	Yes	ISO 8601 timestamp
site.name	Yes	Site name
site.domain	Yes	Domain without protocol
site.description	No	Shown in MCP tool description
site.toolPrefix	No	Tool name prefix (default: website)
pageCount	Yes	Number of pages
pages[].url	Yes	Path starting with /
pages[].title	Yes	Page title
pages[].body	No	Full text (recommended)

---

Development

bun run dev          # Local development server
bun run type-check   # TypeScript checking
bun run lint:fix     # Lint and fix
bun run format       # Format code
bun run deploy       # Deploy to Cloudflare

Note: This is a template repository. The bun run deploy command is for users who clone this template to deploy their own MCP server. To contribute to this template itself, use standard git workflows (git push).

---

Troubleshooting

"Search index not found in R2 bucket"

1. Check the bucket exists: bunx wrangler r2 bucket list
2. Check the file was uploaded: bunx wrangler r2 object list my-site-mcp-data
3. Verify the bucket name in wrangler.jsonc matches

MCP client won't connect

1. Use the /mcp endpoint (recommended) or /sse for legacy clients
2. Visit your worker URL in a browser — you should see JSON
3. Make sure the URL includes https://

Search returns no results

1. Validate your index: bun run validate ./search-index.json
2. Check that pages have body content
3. Try broader search terms

Wrong tool names

Tool names come from toolPrefix in your search index. Regenerate and re-upload the index with the correct value.

Local development

You need a local copy of the search index:

mkdir -p .wrangler/state/r2/my-site-mcp-data
cp search-index.json .wrangler/state/r2/my-site-mcp-data/search-index.json

---

Examples

Two sites using this approach:

REMnux Documentation

MCP server for REMnux, the Linux toolkit for malware analysis. It gives your AI guidance on the REMnux tools, installation steps, and malware analysis workflows.

Repo: github.com/REMnux/remnux-docs-mcp-server

# Claude Code
claude mcp add remnux-docs --transport http https://docs-mcp.remnux.org/mcp --scope user

Lenny Zeltser's Website

MCP server for zeltser.com. It gives your AI guidance on drafting IR reports, shaping cybersecurity product strategy, and other security leadership topics.

# Claude Code
claude mcp add zeltser-search --transport http https://website-mcp.zeltser.com/mcp --scope user

---

AI Agent Quick Reference

Key Files

File	Purpose
src/index.ts	Worker entry point: MCP server, tool definitions, Fuse.js search, R2 index cache
src/types.ts	Shared TypeScript types for the search index schema
wrangler.jsonc	Cloudflare deployment config (Worker name, R2 binding, routes)
adapters/	Index generators for Astro, Hugo, and generic markdown sites
scripts/validate-index.ts	Validates search-index.json against the v3.0 schema

Architecture

Markdown Files → Adapter (build time) → search-index.json → R2 → Worker (MCP) → AI Client

• Adapters run at build time to generate search-index.json
• The Worker loads the index from R2 with 4-hour in-memory caching (uses etag revalidation on expiry)
• Fuse.js provides fuzzy search across titles, abstracts, body text, and topics
• Durable Objects manage persistent MCP client sessions

Common Dev Tasks

bun run dev          # Local dev server (needs local search-index.json in .wrangler/)
bun run deploy       # Deploy Worker to Cloudflare
bun run type-check   # TypeScript checking
bun run validate ./search-index.json  # Validate index

Security Notes

• No authentication: any client with the endpoint URL can query all indexed content
• Designed for public content only
• R2 bucket is private but Worker exposes contents via MCP tools

---

Author

Lenny Zeltser: Builder of security products and programs. Teacher of those who run them.