About Web Page Fetcher
Web Page Fetcher is an MCP server in the Other category: fetch web pages with SSRF protection. Returns clean, readable markdown from any URL.
Install
git clone https://github.com/paimonchan/paimon-mcp-fetchThis server has no ConduID identity, so agent calls to it are not receipted. Pin the version you install and review the source before granting it credentials.
Ask AI
Ask AI about Web Page Fetcher
Powered by Claude · Grounded in docs
Security checks
- ✓README presentThe repository ships documentation.
5 checks have not been run yet.
README
paimon-mcp-fetch
Give your AI assistant the ability to read any webpage.
A lightweight MCP server that fetches URLs and returns clean, readable markdown. Built with Go — starts in 5ms, uses ~10MB RAM, zero runtime dependencies.
What It Does
You give it a URL → it returns clean markdown.
Good for:
- Reading articles, blog posts, documentation
- Extracting data from news sites, forums, schedules
- Summarizing web content for your AI assistant
- Getting structured output (headings, lists, tables, code blocks preserved)
Not good for:
- Scraping login-protected pages
- Bypassing paywalls
- Replacing a full browser automation tool
Quick Start
1. Install
Pick one method:
# Go (recommended)
go install github.com/paimonchan/paimon-mcp-fetch/cmd/paimon-mcp-fetch@latest
# Homebrew (macOS/Linux)
brew tap paimonchan/tap
brew install paimon-mcp-fetch
# Scoop (Windows)
scoop bucket add paimonchan https://github.com/paimonchan/scoop-bucket
scoop install paimon-mcp-fetch
# Winget (Windows)
winget install paimonchan.paimon-mcp-fetch
# Docker
docker run -i --rm ghcr.io/paimonchan/paimon-mcp-fetch:latest
2. Configure Your AI Assistant
Add this to your MCP client config:
{
"mcp": {
"paimon-mcp-fetch": {
"type": "local",
"command": ["paimon-mcp-fetch"],
"enabled": true
}
}
}
3. Done
Your AI can now read any URL you give it.
Why This Over Other Fetch Tools?
| paimon-mcp-fetch | Basic text fetch | |
|---|---|---|
| Output | Structured markdown | Plain text |
| Article extraction | Readability algorithm (strips ads, nav, sidebars) | Raw HTML body |
| Images | Optional extraction + processing | None |
| JS rendering | Optional (headless Chrome) | Static only |
| Caching | Built-in LRU cache | None |
| Rate limiting | Per-domain, configurable | None |
| SSRF protection | 7-layer defense | None |
| Startup time | ~5ms | Varies |
| Memory | ~10MB | Varies |
Configuration
Everything is controlled via environment variables. You probably don't need to change anything — defaults work well for most use cases.
| Variable | Default | What it does |
|---|---|---|
PAIMON_MCP_FETCH_TIMEOUT_MS |
12000 |
Request timeout (ms) |
PAIMON_MCP_FETCH_MAX_HTML_BYTES |
10485760 |
Max page size (10MB) |
PAIMON_MCP_FETCH_CACHE_TTL_SECS |
300 |
Cache lifetime (5 min) |
PAIMON_MCP_FETCH_RATE_LIMIT_PER_SECOND |
5.0 |
Requests/sec per domain |
PAIMON_MCP_FETCH_RATE_LIMIT_BURST |
10 |
Max burst size |
PAIMON_MCP_FETCH_RETRY_MAX_ATTEMPTS |
3 |
Retry on transient errors |
PAIMON_MCP_FETCH_JS_RENDER_ENABLED |
false |
Enable headless Chrome |
Optional Features
Image Processing
Extract and process images from webpages:
go build -tags image -o paimon-mcp-fetch ./cmd/paimon-mcp-fetch/
JS Rendering
For JavaScript-heavy sites (SPAs, dynamic content):
go build -tags jsrender -o paimon-mcp-fetch ./cmd/paimon-mcp-fetch/
PAIMON_MCP_FETCH_JS_RENDER_ENABLED=true ./paimon-mcp-fetch
Note: Requires Chrome or Chromium installed. Slower (~3-5s/page) but handles sites that static fetch can't.
Supported AI Assistants
Works with any MCP-compatible client. Add the configuration below to your client's config file.
OpenCode
Config file: ~/.config/opencode/opencode.json
{
"mcp": {
"paimon-mcp-fetch": {
"type": "local",
"command": ["paimon-mcp-fetch"],
"enabled": true
}
}
}
Claude Desktop
Config file: claude_desktop_config.json
{
"mcpServers": {
"paimon-mcp-fetch": {
"command": "paimon-mcp-fetch"
}
}
}
Cursor
Config file: .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)
{
"mcpServers": {
"paimon-mcp-fetch": {
"command": "paimon-mcp-fetch",
"env": {}
}
}
}
VS Code (GitHub Copilot)
Config file: .vscode/mcp.json (workspace) or user settings
{
"servers": {
"paimon-mcp-fetch": {
"type": "stdio",
"command": "paimon-mcp-fetch"
}
}
}
Cline
Config file: .cline/mcp.json
{
"mcpServers": {
"paimon-mcp-fetch": {
"command": "paimon-mcp-fetch"
}
}
}
Windsurf
Config file: .windsurf/mcp.json
{
"mcpServers": {
"paimon-mcp-fetch": {
"command": "paimon-mcp-fetch"
}
}
}
With Custom Configuration
All clients support environment variables. Example with custom timeout and cache:
{
"mcpServers": {
"paimon-mcp-fetch": {
"command": "paimon-mcp-fetch",
"env": {
"PAIMON_MCP_FETCH_TIMEOUT_MS": "30000",
"PAIMON_MCP_FETCH_CACHE_TTL_SECS": "600"
}
}
}
}
Security
Built with security in mind from day one:
- SSRF Protection — 7-layer defense against server-side request forgery
- Private IP blocking — Can't access localhost or internal networks
- Redirect validation — Every redirect hop re-checked for safety
- Size limits — Stream-based reading, no memory bombs
- Timeouts — All requests have deadlines
- No secrets in logs — API keys and tokens never logged
Windows Antivirus
Some antivirus may flag unsigned Go binaries as a false positive. This is a known industry issue. Solutions:
- Use
go install— antivirus sees the signed Go compiler - Use Docker — no local binary
- Build from source — verify the code yourself
Architecture
Built with Clean Architecture principles:
MCP Server → UseCase → Domain (entities, ports)
↑
Adapters implement interfaces
- Domain — Business rules, zero external dependencies
- UseCase — Orchestration logic
- Adapters — HTTP client, content extractor, cache, rate limiter, image processor, JS renderer
Full details in the project plan.
License
MIT — do whatever you want with it.
Built with Go. Zero runtime dependencies. Single binary. ~10MB RAM. Starts in 5ms.
README copy captured when this server was indexed; it may lag the repository. The original is authoritative.