Fetch URL MCP Server
Fetch public web pages and convert them into clean, AI-readable Markdown.
Overview
Fetch URL is a Model Context Protocol (MCP) server that fetches public web pages, extracts meaningful content using Mozilla's Readability algorithm, and converts the result into clean Markdown optimized for LLM context windows. It handles noise removal, caching, SSRF protection, async task execution, and supports both stdio and Streamable HTTP transports.
[!NOTE]Content extraction quality varies depending on the HTML structure and complexity of the source page. Fetch URL works best with standard article and documentation layouts. Pages relying on client-side JavaScript rendering may yield incomplete results.
Key Features
- HTML to Markdown — Content extraction via Mozilla Readability + node-html-markdown
- Noise removal — Strips navigation, ads, cookie banners, and other non-content elements
- In-memory LRU cache — Faster repeat fetches with configurable TTL (24 h default)
- Raw URL rewriting — Auto-converts GitHub, GitLab, Bitbucket, and Gist URLs to raw content endpoints
Tech Stack
| Component | Technology |
|---|---|
| Runtime | Node.js >= 24 |
| Language | TypeScript 5.9 |
| MCP SDK | @modelcontextprotocol/sdk ^1.26.0 |
| Content Extraction | @mozilla/readability ^0.6.0 |
| DOM Parsing | linkedom ^0.18.12 |
| Markdown Conversion | node-html-markdown ^2.0.0 |
| Schema Validation | zod ^4.3.6 |
| Package Manager | npm |
Architecture
URL → Validate → DNS Preflight → HTTP Fetch → Decompress
→ Truncate HTML → Readability Extract → Noise Removal
→ Markdown Convert → Cleanup Pipeline → Cache → Response
- URL Validation — Normalize, block private hosts, transform raw-content URLs (GitHub, GitLab, Bitbucket)
- Fetch — HTTP request with redirect following, DNS preflight SSRF checks, and size limits (10 MB)
- Transform — Offloaded to worker threads: parse HTML with
linkedom, extract with Readability, remove DOM noise, convert to Markdown - Cleanup — Multi-pass Markdown normalization (heading promotion, spacing, skip-link removal)
- Cache + Respond — Store result in LRU cache, apply inline content limits, return structured content
Repository Structure
fetch-url-mcp/
├── assets/ # Server icon (logo.svg)
├── scripts/ # Build & test orchestration
├── src/
│ ├── workers/ # Worker-thread child for HTML transforms
│ ├── index.ts # CLI entrypoint, transport wiring, shutdown
│ ├── server.ts # McpServer lifecycle and registration
│ ├── tools.ts # fetch-url tool definition and pipeline
│ ├── fetch.ts # URL normalization, SSRF, HTTP fetch
│ ├── transform.ts # HTML-to-Markdown pipeline, worker pool
│ ├── config.ts # Env-driven configuration
│ ├── resources.ts # MCP resource/template registration
│ ├── prompts.ts # MCP prompt registration (get-help)
│ ├── mcp.ts # Task execution management
│ ├── http-native.ts # Streamable HTTP server, auth, sessions
│ └── instructions.md # Server instructions embedded at runtime
├── tests/ # Unit/integration tests (Node.js test runner)
├── package.json
├── tsconfig.json
└── AGENTS.md
Requirements
- Node.js >= 24
Quickstart
npx -y @j0hanz/fetch-url-mcp@latest --stdio
Add to your MCP client configuration:
{
"mcpServers": {
"fetch-url-mcp": {
"command": "npx",
"args": ["-y", "@j0hanz/fetch-url-mcp@latest", "--stdio"]
}
}
}
Installation
NPX (Recommended)
No installation required — runs directly:
npx -y @j0hanz/fetch-url-mcp@latest --stdio
Global Install
npm install -g @j0hanz/fetch-url-mcp
fetch-url-mcp --stdio
From Source
git clone https://github.com/j0hanz/fetch-url-mcp.git
cd fetch-url-mcp
npm install
npm run build
node dist/index.js --stdio
Docker
docker compose up --build
Configuration
Runtime Modes
| Flag | Description |
|---|---|
--stdio, -s |
Run in stdio mode (for desktop MCP clients) |
--help, -h |
Show usage help |
--version, -v |
Print server version |
When no --stdio flag is passed, the server starts in HTTP mode (Streamable HTTP on port 3000 by default).
Environment Variables
Core Settings
| Variable | Default | Description |
|---|---|---|
HOST |
127.0.0.1 |
HTTP server bind address |
PORT |
3000 |
HTTP server port (1024–65535) |
LOG_LEVEL |
info |
Log level: debug, info, warn, error |
FETCH_TIMEOUT_MS |
15000 |
HTTP fetch timeout in ms (1000–60000) |
CACHE_ENABLED |
true |
Enable/disable in-memory content cache |
USER_AGENT |
fetch-url-mcp/{version} |
Custom User-Agent header |
ALLOW_REMOTE |
false |
Allow remote connections in HTTP mode |
ALLOWED_HOSTS |
(empty) | Comma-separated host/origin allowlist for HTTP mode |
Task Management
| Variable | Default | Description |
|---|---|---|
TASKS_MAX_TOTAL |
5000 |
Maximum retained task records across all owners |
TASKS_MAX_PER_OWNER |
1000 |
Maximum retained task records per session/client |
Authentication (HTTP Mode)
| Variable | Default | Description |
|---|---|---|
ACCESS_TOKENS |
(empty) | Comma-separated static bearer tokens |
API_KEY |
(empty) | Single API key (added to static tokens) |
OAUTH_ISSUER_URL |
(empty) | OAuth issuer URL (enables OAuth mode) |
OAUTH_AUTHORIZATION_URL |
(empty) | OAuth authorization endpoint |
OAUTH_TOKEN_URL |
(empty) | OAuth token endpoint |
OAUTH_INTROSPECTION_URL |
(empty) | OAuth token introspection endpoint |
OAUTH_REVOCATION_URL |
(empty) | OAuth token revocation endpoint |
OAUTH_REGISTRATION_URL |
(empty) | OAuth dynamic client registration |
OAUTH_REQUIRED_SCOPES |
(empty) | Required OAuth scopes |
OAUTH_CLIENT_ID |
(empty) | OAuth client ID |
OAUTH_CLIENT_SECRET |
(empty) | OAuth client secret |
Transform & Workers
| Variable | Default | Description |
|---|---|---|
TRANSFORM_WORKER_MODE |
threads |
Worker mode: threads or process |
TRANSFORM_WORKER_MAX_OLD_GENERATION_MB |
(unset) | V8 old generation heap limit per worker |
TRANSFORM_WORKER_MAX_YOUNG_GENERATION_MB |
(unset) | V8 young generation heap limit per worker |
TRANSFORM_WORKER_CODE_RANGE_MB |
(unset) | V8 code range limit per worker |
TRANSFORM_WORKER_STACK_MB |
(unset) | Stack size limit per worker |
Content Tuning
| Variable | Default | Description |
|---|---|---|
MAX_INLINE_CONTENT_CHARS |
0 |
Global inline markdown limit (0 = unlimited) |
FETCH_URL_MCP_EXTRA_NOISE_TOKENS |
(empty) | Additional CSS class/id tokens for noise removal |
FETCH_URL_MCP_EXTRA_NOISE_SELECTORS |
(empty) | Additional CSS selectors for noise removal |
MARKDOWN_HEADING_KEYWORDS |
(built-in list) | Keywords triggering heading promotion |
FETCH_URL_MCP_LOCALE |
(system) | Locale for content processing |
Server Tuning
| Variable | Default | Description |
|---|---|---|
SERVER_MAX_CONNECTIONS |
0 (unlimited) |
Maximum concurrent HTTP connections |
SERVER_BLOCK_PRIVATE_CONNECTIONS |
false |
Block connections from private IP ranges |
Hardcoded Defaults
| Setting | Value |
|---|---|
| Max HTML size | 10 MB |
| Max inline content chars | 0 (unlimited, configurable) |
| Fetch timeout | 15 s |
| Transform timeout | 30 s |
| Tool timeout | Fetch + Transform + 5 s padding |
| Max redirects | 5 |
| Cache TTL | 86400 s (24 h) |
| Cache max keys | 100 |
| Rate limit | 100 requests / 60 s |
| Max sessions | 200 |
| Session TTL | 30 min |
| Max URL length | 2048 chars |
| Worker pool max scale | 4 |
Usage
Stdio Mode
fetch-url-mcp --stdio
The server communicates via JSON-RPC over stdin/stdout. All MCP clients that support stdio transport can connect directly.
HTTP Mode
fetch-url-mcp
# or
PORT=8080 HOST=0.0.0.0 ALLOW_REMOTE=true fetch-url-mcp
The server starts a Streamable HTTP endpoint at /mcp. Authenticate with bearer tokens via the ACCESS_TOKENS or API_KEY environment variables.
For POST /mcp, clients should send:
Accept: application/json, text/event-streamMCP-Protocol-Version: 2025-11-25(or2025-03-26for legacy clients)
MCP Surface
Tools
fetch-url
Fetches a webpage and converts it to clean Markdown format optimized for LLM context.
Useful for:
- Reading documentation, blog posts, or articles
- Extracting main content while removing navigation and ads
- Caching content to speed up repeated queries
Limitations:
- Does not execute complex client-side JavaScript interactions
- Inline output may be truncated when
MAX_INLINE_CONTENT_CHARSis set
Parameters
| Parameter | Type | Required | Default | Description |
|---|---|---|---|---|
url |
string (URL) |
Yes | — | The URL of the webpage to fetch (http/https, max 2048 chars) |
skipNoiseRemoval |
boolean |
No | false |
Preserve navigation, footers, and other elements normally filtered |
forceRefresh |
boolean |
No | false |
Bypass cache and fetch fresh content |
maxInlineChars |
number |
No | 0 |
Per-call inline markdown limit (0 = unlimited; global cap still applies) |
Returns
{
"url": "https://example.com",
"inputUrl": "https://example.com",
"resolvedUrl": "https://example.com",
"finalUrl": "https://example.com",
"title": "Example Domain",
"metadata": {
"title": "Example Domain",
"description": "...",
"author": "...",
"image": "...",
"favicon": "...",
"publishedAt": "...",
"modifiedAt": "..."
},
"markdown": "# Example Domain\n\nThis domain is for use in illustrative examples...",
"fromCache": false,
"fetchedAt": "2026-02-11T12:00:00.000Z",
"contentSize": 1234,
"truncated": false
}
| Field | Type | Description |
|---|---|---|
url |
string |
The canonical URL (pre-raw-transform) |
inputUrl |
string? |
The original URL provided by the caller |
resolvedUrl |
string? |
The normalized/transformed URL that was fetched |
finalUrl |
string? |
Final response URL after redirects |
title |
string? |
Extracted page title |
metadata |
object? |
Extracted metadata (title, description, author, image, favicon, publishedAt, modifiedAt) |
markdown |
string? |
Extracted content in Markdown format |
fromCache |
boolean? |
Whether the response was served from cache |
fetchedAt |
string? |
ISO timestamp for fetch/cache retrieval |
contentSize |
number? |
Full markdown size before inline truncation |
truncated |
boolean? |
Whether inline markdown was truncated |
error |
string? |
Error message if the request failed |
statusCode |
number? |
HTTP status code for failed requests |
details |
object? |
Additional error details |
Annotations
| Annotation | Value |
|---|---|
readOnlyHint |
true |
destructiveHint |
false |
idempotentHint |
true |
openWorldHint |
true |
Async Task Execution
The fetch-url tool supports optional async task execution (execution.taskSupport: "optional"). Include a task field in the tool call to run the fetch in the background:
{
"method": "tools/call",
"params": {
"name": "fetch-url",
"arguments": { "url": "https://example.com" },
"task": { "ttl": 30000 }
}
}
Then poll tasks/get until the task status is completed or failed, and retrieve the result via tasks/result.
Prompts
| Name | Description |
|---|---|
get-help |
Returns server usage instructions |
Resources
| URI Pattern | MIME Type | Description |
|---|---|---|
internal://instructions |
text/markdown |
Server instructions and usage guidance |
internal://cache/{namespace}/{hash} |
text/markdown |
Cached markdown entries from prior fetch-url calls |
Tasks
The server declares full MCP task support:
| Endpoint | Description |
|---|---|
tasks/list |
List tasks (scoped to session/owner) |
tasks/get |
Get task status by ID |
tasks/result |
Retrieve completed task result |
tasks/cancel |
Cancel an in-flight task |
HTTP Mode Endpoints
| Method | Path | Auth | Description |
|---|---|---|---|
GET |
/health |
No | Health check (minimal payload) |
GET |
/health?verbose=true |
Yes* | Detailed diagnostics and runtime metrics |
POST |
/mcp |
Yes | MCP JSON-RPC (Streamable HTTP) |
GET |
/mcp |
Yes | SSE stream for server-initiated messages |
DELETE |
/mcp |
Yes | Terminate MCP session |
GET |
/mcp/downloads/{namespace}/{hash} |
Yes | Download cached content |
* verbose=true can be read without auth only for local-only deployments (ALLOW_REMOTE=false).
Session Behavior
- Sessions are created on the first
POST /mcprequest with aninitializemessage - Session ID is returned in the
mcp-session-idresponse header - Sessions expire after 30 minutes of inactivity (max 200 concurrent)
Authentication
- Static tokens: Set
ACCESS_TOKENSorAPI_KEYenvironment variables; pass asAuthorization: Bearer <token> - OAuth: Configure
OAUTH_*environment variables to enable OAuth 2.0 token introspection
Client Configuration Examples
VS Code / VS Code InsidersAdd to your VS Code settings (.vscode/mcp.json or User Settings):
{
"servers": {
"fetch-url-mcp": {
"command": "npx",
"args": ["-y", "@j0hanz/fetch-url-mcp@latest", "--stdio"]
}
}
}
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"fetch-url-mcp": {
"command": "npx",
"args": ["-y", "@j0hanz/fetch-url-mcp@latest", "--stdio"]
}
}
}
Cursor
Or manually add to Cursor MCP settings:
{
"mcpServers": {
"fetch-url-mcp": {
"command": "npx",
"args": ["-y", "@j0hanz/fetch-url-mcp@latest", "--stdio"]
}
}
}
Windsurf
Add to your Windsurf MCP configuration:
{
"mcpServers": {
"fetch-url-mcp": {
"command": "npx",
"args": ["-y", "@j0hanz/fetch-url-mcp@latest", "--stdio"]
}
}
}
Docker
Use the published image from GitHub Container Registry:
{
"mcpServers": {
"fetch-url-mcp": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"ghcr.io/j0hanz/fetch-url-mcp:latest",
"--stdio"
]
}
}
}
Or build and run locally:
docker build -t fetch-url-mcp .
docker run -i --rm fetch-url-mcp --stdio
Security
SSRF Protection
Fetch URL blocks requests to private and internal network addresses:
- Blocked hosts:
localhost,127.0.0.0/8,10.0.0.0/8,172.16.0.0/12,192.168.0.0/16,169.254.0.0/16,100.64.0.0/10 - Blocked IPv6:
::1,fc00::/7,fe80::/10, IPv4-mapped private addresses - Cloud metadata:
169.254.169.254(AWS),metadata.google.internal,metadata.azure.com,100.100.100.200(Azure IMDS)
DNS preflight checks run on every redirect hop to prevent DNS rebinding attacks.
Stdio Transport Safety
The server never writes non-protocol data to stdout. All logs and diagnostics go to stderr.
Rate Limiting
HTTP mode enforces a rate limit of 100 requests per 60-second window per client.
Content Safety
- HTML downloads are capped at 10 MB
- Worker threads run in isolation with configurable resource limits
- Auth tokens are stored in-memory only and compared using timing-safe equality
Development Workflow
Install Dependencies
npm install
Scripts
| Script | Command | Description |
|---|---|---|
dev |
npm run dev |
TypeScript watch mode |
dev:run |
npm run dev:run |
Run compiled output with watch + .env |
build |
npm run build |
Clean, compile, copy assets, make executable |
start |
npm start |
Run compiled server |
test |
npm test |
Run test suite (Node.js native test runner) |
test:coverage |
npm run test:coverage |
Run tests with coverage |
lint |
npm run lint |
ESLint |
lint:fix |
npm run lint:fix |
ESLint with auto-fix |
format |
npm run format |
Prettier |
type-check |
npm run type-check |
TypeScript type checking |
inspector |
npm run inspector |
Build and launch MCP Inspector |
Build and Release
npm run build # Clean → Compile → Copy Assets → chmod
npm run prepublishOnly # Lint → Type-Check → Build
npm publish # Publish to npm
CI/CD is handled via a GitHub Actions workflow (release.yml) that runs lint, type-check, test, build, and publishes to npm with version bumping.
Troubleshooting
MCP Inspector
Use the built-in inspector to test the server interactively:
npm run inspector
Common Issues
| Issue | Solution |
|---|---|
VALIDATION_ERROR on URL |
URL is blocked (private IP/localhost) or malformed. Do not retry. |
queue_full error |
Worker pool busy. Wait briefly, then retry or use async task mode. |
| Garbled output | Binary content (images, PDFs) cannot be converted. Ensure the URL serves HTML. |
| No output in stdio mode | Ensure --stdio flag is passed. Without it, the server starts in HTTP mode. |
| Auth errors in HTTP mode | Set ACCESS_TOKENS or API_KEY env var and pass as Authorization: Bearer <token>. |
Stdout / Stderr Guidance
In stdio mode, stdout is reserved exclusively for MCP JSON-RPC messages. Logs and diagnostics are written to stderr. Never pipe stdout to a log file when using stdio transport.