firecrawl-lite-mcp-server
Enables privacy-first web scraping and structured data extraction using a local headless browser and your own LLM key. Supports tools for scraping, batch scraping, data extraction with prompts or schemas, and screenshots.
README
<div align="center">
<img src="docs/banner.jpg" alt="Firecrawl Lite MCP Server" width="100%" />
Firecrawl Lite MCP Server
Privacy-first web scraping and data extraction for any MCP client — powered by local browser automation and your own LLM key.
</div>
Firecrawl Lite is a standalone Model Context Protocol server that gives any MCP client — Claude Desktop, Claude Code, Cursor, and others — the ability to scrape web pages and extract structured data. Pages are fetched and rendered with a local, stealth-enabled headless browser, and structured extraction is performed by your own LLM provider. There is no Firecrawl account and no third-party scraping service in the loop: the only API key you bring is the one for the LLM you already use.
Contents
- Why Firecrawl Lite
- Available Tools
- Quick Start
- Configuration
- Remote Deployment
- Advanced Configuration
- Usage Examples
- Troubleshooting
- Container Images
- Development
- Credits
- License
Why Firecrawl Lite
Privacy-first. Scraping and rendering happen on your own machine or server. Page content is only ever sent to the LLM provider you explicitly configure — nothing is routed through a third-party scraping cloud.
Bring your own model. Works with any OpenAI-compatible chat/completions endpoint: OpenAI, xAI (Grok), Anthropic, OpenRouter, Synthetic, or a local model via Ollama. You pay only for the LLM tokens you use.
Lightweight and self-contained. A single Node.js process with a bundled headless browser. Run it locally over stdio, or deploy it as one container behind HTTP/SSE. Multi-architecture images (amd64 / arm64) are published on every release.
Built for real scraping. Stealth browser automation, rotating user agents, configurable delays, optional upstream proxies (including port-range rotation), and tunable retry/backoff.
Available Tools
| Tool | Description | Required params | Optional params |
|---|---|---|---|
scrape_page |
Fetch and render a single page, returning clean text/markdown. | url |
onlyMainContent |
batch_scrape |
Scrape multiple URLs in one request (up to 10). | urls[] |
onlyMainContent |
extract_data |
Extract structured data from pages using a natural-language prompt and your LLM. | urls[], prompt |
enableWebSearch |
extract_with_schema |
Extract data conforming to a supplied JSON Schema. | urls[], schema |
prompt, enableWebSearch |
screenshot |
Capture a screenshot of a page via the stealth browser. | url |
width, height, fullPage |
Quick Start
The fastest way to use Firecrawl Lite locally is over stdio via npx — no install or container required. Add the LLM credentials for the provider of your choice (see LLM provider examples).
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"firecrawl-lite": {
"command": "npx",
"args": ["-y", "@ariangibson/firecrawl-lite-mcp-server"],
"env": {
"LLM_API_KEY": "your_llm_api_key_here",
"LLM_PROVIDER_BASE_URL": "https://api.openai.com/v1",
"LLM_MODEL": "gpt-5.5"
}
}
}
}
Claude Code (CLI)
claude mcp add firecrawl-lite npx -- -y @ariangibson/firecrawl-lite-mcp-server \
--env LLM_API_KEY=your_key \
--env LLM_PROVIDER_BASE_URL=https://api.openai.com/v1 \
--env LLM_MODEL=gpt-5.5
Cursor
Add to your Cursor MCP configuration (~/.cursor/mcp.json):
{
"mcpServers": {
"firecrawl-lite": {
"command": "npx",
"args": ["-y", "@ariangibson/firecrawl-lite-mcp-server"],
"env": {
"LLM_API_KEY": "your_llm_api_key_here",
"LLM_PROVIDER_BASE_URL": "https://api.openai.com/v1",
"LLM_MODEL": "gpt-5.5"
}
}
}
}
Configuration
All configuration is via environment variables. Only the three LLM variables are required; everything else has sensible defaults.
Required
| Variable | Description |
|---|---|
LLM_API_KEY |
API key for your LLM provider. |
LLM_PROVIDER_BASE_URL |
Base URL of an OpenAI-compatible API (the server calls {base_url}/chat/completions). |
LLM_MODEL |
Model name to use for extraction. |
Optional LLM tuning
These are passed straight through to the provider's chat/completions request. Leave any of them unset to use the default; the optional sampling parameters are omitted from the request entirely when unset.
| Variable | Default | Notes |
|---|---|---|
LLM_TEMPERATURE |
0.1 |
Sampling temperature. |
LLM_MAX_TOKENS |
2000 |
Maximum tokens in the response. Raise this if extractions are being truncated. |
LLM_TOP_P |
unset | Nucleus sampling; omitted from the request unless set. |
LLM_REASONING_EFFORT |
unset | reasoning_effort for reasoning-capable models; omitted unless set. |
LLM provider examples
# OpenAI
LLM_PROVIDER_BASE_URL=https://api.openai.com/v1
LLM_MODEL=gpt-5.5
# xAI (Grok)
LLM_PROVIDER_BASE_URL=https://api.x.ai/v1
LLM_MODEL=grok-4
# Anthropic
LLM_PROVIDER_BASE_URL=https://api.anthropic.com/v1
LLM_MODEL=claude-haiku-4-5
# OpenRouter
LLM_PROVIDER_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL=openai/gpt-5.5
# Local (Ollama)
LLM_PROVIDER_BASE_URL=http://localhost:11434/v1
LLM_MODEL=llama3.3
See .env.example for the full, annotated list of variables.
Remote Deployment
For remote or containerized use, enable at least one HTTP transport — both are disabled by default for security. Choose based on your client:
ENABLE_HTTP_STREAMABLE_ENDPOINT=true→ exposes/mcp(Claude Code, modern remote MCP clients).ENABLE_SSE_ENDPOINT=true→ exposes/sse(Claude Desktop viamcp-proxy).
A /health endpoint is always available for health checks.
Docker
docker run -d \
-p 3000:3000 \
-e ENABLE_HTTP_STREAMABLE_ENDPOINT=true \
-e LLM_API_KEY=your_key_here \
-e LLM_PROVIDER_BASE_URL=https://api.openai.com/v1 \
-e LLM_MODEL=gpt-5.5 \
ariangibson/firecrawl-lite-mcp-server:latest
Docker Compose / Portainer / Swarm
A ready-to-use docker-compose.yml is included. Set your variables in a .env file and deploy:
docker compose up -d
Note for Docker Swarm / Portainer: the published image is
node:20-alpine, which does not includecurl. The bundled compose file uses awget-based health check for this reason — see Troubleshooting if you have customized it.
Remote client configuration
Claude Code (Streamable HTTP):
claude mcp add firecrawl-lite-remote http://your-server:3000/mcp -t http
Claude Desktop — Connectors (recommended, HTTPS only):
Settings → Connectors → add https://your-server.com:3000/mcp. Requires a valid TLS certificate.
Claude Desktop — mcp-proxy (HTTP fallback, no certificate):
pip install mcp-proxy
{
"mcpServers": {
"firecrawl-lite": {
"command": "mcp-proxy",
"args": ["http://your-server:3000/sse"]
}
}
}
Advanced Configuration
Proxy
Route the scraping browser through an upstream proxy. A port range (e.g. :10001-10010) enables automatic rotation across ports.
PROXY_SERVER_URL=http://proxy.example.com:10001-10010
PROXY_SERVER_USERNAME=your-username
PROXY_SERVER_PASSWORD=your-password
By default the proxy is used only for scraping target sites — LLM provider API calls go out directly. Routing your own LLM calls through a rotating (often residential) proxy is slower, can trip provider abuse detection, and may fail TLS. If you specifically need the LLM call proxied as well, opt in with:
PROXY_LLM_API=true
Anti-detection and rate limiting
SCRAPE_USER_AGENT accepts either a single string or a JSON array of strings to rotate through. When using a JSON array, keep it on a single line.
SCRAPE_USER_AGENT=["Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) ... Safari/537.36","Mozilla/5.0 (Windows NT 10.0; Win64; x64) ... Safari/537.36"]
SCRAPE_VIEWPORT_WIDTH=1920
SCRAPE_VIEWPORT_HEIGHT=1080
SCRAPE_DELAY_MIN=1000 # min delay before navigation (ms)
SCRAPE_DELAY_MAX=3000 # max delay before navigation (ms)
SCRAPE_BATCH_DELAY_MIN=2000 # min delay between batch requests (ms)
SCRAPE_BATCH_DELAY_MAX=5000 # max delay between batch requests (ms)
SCRAPE_SETTLE_MAX_MS=3000 # max wait for the DOM to stop changing after load (ms); raise for slow JS sites
After load the scraper scrolls to trigger lazy/AJAX content and waits for the
DOM to settle (exiting early once stable). Pages that inject content via a long
setTimeout may need a higher SCRAPE_SETTLE_MAX_MS.
Retry and backoff
FIRECRAWL_RETRY_MAX_ATTEMPTS=3
FIRECRAWL_RETRY_INITIAL_DELAY=1000
FIRECRAWL_RETRY_MAX_DELAY=10000
FIRECRAWL_RETRY_BACKOFF_FACTOR=2
Usage Examples
Scrape a page
{ "name": "scrape_page", "arguments": { "url": "https://example.com" } }
Batch scrape
{
"name": "batch_scrape",
"arguments": {
"urls": ["https://example.com", "https://example.org"],
"onlyMainContent": true
}
}
Extract with a prompt
{
"name": "extract_data",
"arguments": {
"urls": ["https://example.com"],
"prompt": "Extract the main article title and a one-sentence summary."
}
}
Extract with a JSON Schema
{
"name": "extract_with_schema",
"arguments": {
"urls": ["https://example.com"],
"schema": {
"type": "object",
"properties": {
"title": { "type": "string" },
"description": { "type": "string" }
}
}
}
}
Troubleshooting
Chrome / Chromium issues
The local image bundles Chromium. For the npx install, Chrome is downloaded on first use. If it fails:
npx puppeteer browsers install chrome
# or reset a corrupted install
rm -rf ~/.cache/puppeteer && npx puppeteer browsers install chrome
Extraction returns an error
scrape_page working but extract_data failing points to the LLM call rather than scraping. The server logs the upstream status, error code, and response body to stderr (LLM extract_data request failed: ...) and surfaces the HTTP status in the tool result. Common causes:
- HTTP 401 — invalid
LLM_API_KEY. - HTTP 400 — wrong
LLM_MODEL, or a tuning parameter the model rejects (e.g.LLM_MAX_TOKENSabove the model's limit, orLLM_REASONING_EFFORTon a non-reasoning model). - HTTP 429 — provider rate limit.
Container keeps restarting or is killed with SIGTERM
If the logs show the server start (listening on port 3000) and then exit with npm error signal SIGTERM, the container is being killed by a failing health check, not by the app. The node:20-alpine image does not include curl, so a curl-based health check always fails and Swarm restarts the task in a loop. Use a wget-based check (busybox provides wget):
healthcheck:
test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://localhost:3000/health || exit 1"]
interval: 30s
timeout: 10s
retries: 3
start_period: 40s
The bundled docker-compose.yml already uses this form.
Container Images
Pre-built, multi-architecture (amd64, arm64) images are published automatically on every push to main and on release:
- Docker Hub:
ariangibson/firecrawl-lite-mcp-server:latest - GitHub Container Registry:
ghcr.io/ariangibson/firecrawl-lite-mcp-server:latest
Development
npm install # install dependencies
npm run build # compile TypeScript to dist/
npm run lint # type-check without emitting
npm test # run the unit test suite
npm start # run the built server
Unit tests cover the pure helpers in src/utils.ts (URL validation, proxy/user-agent parsing, LLM request construction) and src/htmlToMarkdown.ts (HTML cleaning and Markdown conversion), and run in CI against Node 18, 20, and 22.
Credits
Inspired by the excellent work of the Firecrawl team at Mendable.ai and their official Firecrawl MCP Server. Firecrawl Lite is an independent, self-hosted take on the same idea — huge thanks to them for pioneering web scraping for the MCP ecosystem.
Looking for a fully managed, enterprise-grade scraping platform? Check out firecrawl.com.
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.