websift

A lightweight, free, self-hosted MCP (Model Context Protocol) server that gives AI agents real-time web access — DuckDuckGo search + web page fetching (HTML → Markdown, PDF → text) — with built-in SSRF protection and DNS pinning. No API key required.

Table of Contents

• What It Does
• Why Use This
• Comparison with Alternatives
• Architecture
• Installation
  • Option 1: PyPI (Recommended)
  • Option 2: Docker Compose
  • Option 3: Docker (Manual)
  • Option 4: Local Python (No Docker)
• Usage
  • CLI
  • As a Python Library (Direct Import)
  • As an MCP Server (For AI Clients)
• Tools
• Configuration
• Connecting to AI Clients
  • VS Code (GitHub Copilot)
  • Claude Desktop
  • Claude Code
  • Copilot CLI
  • Cursor
  • Windsurf (Codeium)
  • JetBrains IDEs
  • Any MCP Client (Generic)
• Use Cases
• Security
• Development
• FAQ

What It Does

This MCP server exposes two tools to any AI agent or LLM client:

That's it — simple, focused, and reliable.

Why Use This

Core Strengths

• 🆓 Completely Free (default) — Default DDGS provider needs no API key or subscription. Upstream DuckDuckGo and target sites may still rate-limit or block clients.
• 🪶 Lightweight — Single Python process, ~4 dependencies, runs in a tiny Docker container (python:3.12-slim ≈ 150 MB).
• 🔒 Secure by Default — SSRF protection (global-only IP policy, multi-answer DNS validation, no URL userinfo), DNS pinning + SNI, redirect re-validation, body/decompress limits, and content-type checks.
• 🌐 Universal MCP Compatibility — Works with any MCP client (VS Code, Claude, Cursor, Windsurf, JetBrains, custom agents, etc.).
• 📄 Smart Content Extraction — HTML → Markdown via BeautifulSoup (main-content selection), PDF → text via pypdf only, binary detection, charset order BOM → HTTP → meta → UTF-8.
• 🐙 GitHub README Shortcut — Fetching a github.com/owner/repo URL uses the GitHub API for the raw README (non-credential headers only).
• 🏠 Self-Hosted — You run the process; still makes outbound requests to the search provider and fetched URLs.

Ideal For

• 🐍 Python Scripts & Apps — Import WebSearchClient directly in your code. No server, no Docker, no MCP overhead.
• AI Agents & Agentic Workflows — Give any autonomous agent the ability to search the web and read pages on demand.
• Development Assistants — Let Copilot, Claude, or Cursor look up documentation, error messages, or package info in real time.
• Research & Analysis — Fetch and summarize articles, papers, or documentation pages.
• Cost-Sensitive Deployments — Replace paid web-search APIs (Tavily, Firecrawl, Exa, etc.) with a free self-hosted alternative.
• Air-Gapped / Private Networks — Run entirely offline (search requires internet, but fetch can work with internal URLs if you adjust security rules).

Comparison with Alternatives

When to Choose What

Architecture

┌─────────────┐     MCP Protocol      ┌──────────────────┐
│  AI Client  │ ◄── (streamable-HTTP) ┤   MCP Server     │
│  (Copilot,  │                       │   (FastMCP)      │
│   Claude,   │                       └─────────┬────────┘
│   Cursor…)  │                                 │
└─────────────┘                    ┌────────────┴─────────┐
                                   │  WebSearchClient     │
┌────────────┐                     │  + WorkLimits        │
│  search()  │──► SearchProvider   │                      │
│            │    (default: DDGS)  │                      │
│  fetch()   │──► urllib + SSRF    │                      │
│            │    ├── html.py      │                      │
│            │    ├── http.py      │                      │
│            │    ├── security.py  │                      │
│            │    └── content.py   │                      │
└────────────┘                     └──────────────────────┘

Outbound data flow: search → configured provider (default DuckDuckGo via ddgs); fetch → primary provider (BaseProvider.fetch, generic SSRF-safe by default). With SEARCH_PROVIDER=tavily|exa and PROVIDER_NATIVE_FETCH=true (default), web_fetch may call Tavily /extract or Exa /contents (credits apply) before falling back to generic fetch. Provider API secrets never ride the target page-fetch path; native extract sends only the URL as JSON to the provider origin.

Module Structure

websift/
├── __init__.py       # WebSearchClient, AppSettings, __version__
├── __main__.py       # python -m websift
├── cli.py            # argparse CLI (serve / search / fetch)
├── settings.py       # AppSettings.from_env() — no env read on import
├── concurrency.py    # WorkLimits (search/fetch/PDF bounds)
├── models.py         # SearchResponse / FetchResult internals
├── config.py         # size limits, user-agents, MIME types
├── security.py       # SSRF: global-only IPs, multi-answer DNS, no userinfo
├── content.py        # content-type detection (PDF, binary, HTML heuristics)
├── http.py           # page fetch: redirects, DNS pin + SNI, body/decompress caps
├── html.py           # HTML → Markdown, main content, truncation
├── client.py         # public search/fetch façade
├── provider_http.py  # credentialed provider transport (isolated from page fetch)
├── providers/        # SearchProvider contract, registry, DDGS + others
└── server.py         # create_server / ServerApp
server.py             # thin entry → websift.cli:main

Installation

Option 1: PyPI (Recommended)

pip install websift

That's it — you can now use it as a Python library (direct import) or as an MCP server (for AI clients).

Option 2: Docker Compose

# Clone and start (no .env file required)
git clone <repo-url>
cd websift
docker compose up -d --build

The MCP server will be available at http://localhost:8787/mcp.

• Runs as non-root (uid 10001), entrypoint websift from the installed wheel.
• Inject secrets at runtime only, e.g. MCP_BEARER_TOKEN=... BRAVE_API_KEY=... docker compose up -d or docker compose --env-file .env up -d.
• Image default bind is 0.0.0.0 inside the container network so published ports work; protect with host firewall, reverse proxy, and/or MCP_AUTH_MODE=bearer.

Optional resource limits (compose / orchestrator): ~512 MB RAM and 1 CPU are enough for light use.

Option 3: Docker (Manual)

docker build -t websift .
docker run -d --name websift -p 8787:8787 \
  -e MCP_AUTH_MODE=none \
  websift
# With bearer auth:
# docker run -d --name websift -p 8787:8787 \
#   -e MCP_AUTH_MODE=bearer -e MCP_BEARER_TOKEN='…' websift

Option 4: Local Python (No Docker)

# Recommended: install the package (editable for development)
pip install -e ".[dev]"

# Or runtime deps only (mirrors pyproject.toml)
pip install -r requirements.txt
pip install -e .

# Run the server (console entry or module)
websift serve
# websift                 # same as serve when no command is given
# python -m websift serve
# python server.py serve

Usage

CLI

After pip install websift, the websift command supports help, version, and subcommands:

websift --help
websift --version

# Start MCP server (default when no command is given)
websift
websift serve
websift serve --host 0.0.0.0 --port 9000 --transport streamable-http
websift serve --auth-mode bearer --bearer-token 'a-long-random-secret'
websift serve --provider ddgs --max-results 8 --log-level DEBUG

# One-shot library-style commands (no MCP server)
websift search "Python 3.12 features"
websift search "asyncio tutorial" -n 10 --provider ddgs
websift fetch https://docs.python.org/3/
websift fetch https://example.com/doc.pdf --max-chars 20000

CLI flags override the corresponding environment variables for that process. Full env matrix is under Configuration.

As a Python Library (Direct Import)

Use WebSearchClient directly in your Python code — no server needed:

from websift import WebSearchClient

client = WebSearchClient()

# Search the web (DuckDuckGo by default)
results = client.search("Python 3.12 features")
print(results)

# Fetch a web page (HTML → Markdown, PDF → text)
content = client.fetch("https://docs.python.org/3/")
print(content)

Customizing WebSearchClient

Pass constructor kwargs — no need to edit env vars for library use:

from websift import WebSearchClient, AppSettings
from websift.settings import ProviderSettings, ExtractionSettings, FetchSettings

# Simple tuning
client = WebSearchClient(
    max_results=10,
    timeout=20,              # shared search+fetch timeout (legacy)
    max_page_chars=50_000,
)

# Separate timeouts + provider name + extraction flags
client = WebSearchClient(
    max_results=8,
    search_timeout=15,
    fetch_timeout=45,
    max_page_chars=64_000,
    provider="ddgs",           # or "brave" / "tavily" / "exa" / "searxng"
    include_links=True,
    include_images=False,
    output_format="markdown",  # or "text"
    native_fetch=True,         # Tavily/Exa paid extract when keyed
)

# Keyed provider (API key stays in your process — never accepted via MCP tools)
client = WebSearchClient(
    provider="brave",
    api_key="BSA...",
    # base_url="https://api.search.brave.com",  # optional override
    fallback_providers=["ddgs"],
    max_results=5,
)

# Full settings tree (also: AppSettings.from_env())
settings = AppSettings(
    provider=ProviderSettings(name="ddgs", max_results=10, timeout_seconds=20),
    fetch=FetchSettings(timeout_seconds=45),
    extraction=ExtractionSettings(max_page_chars=50_000, include_links=True),
)
client = WebSearchClient(settings=settings)

# Or load env, then use as-is
client = WebSearchClient(settings=AppSettings.from_env())

Perfect for:

• Custom scripts & automation
• Embedding in your own applications
• Data pipelines & ETL workflows
• Testing & prototyping

As an MCP Server (For AI Clients)

Run the server to expose tools to any MCP-compatible AI client:

# Start the server (default: 127.0.0.1:8787, streamable-http)
websift serve

# Custom bind / transport via CLI (overrides env for this process)
websift serve --host 0.0.0.0 --port 9000 --transport sse

# Or via environment variables
MCP_PORT=9000 MCP_TRANSPORT=sse websift serve

Or via Python:

from websift.server import create_server
from websift.settings import AppSettings

create_server(AppSettings.from_env()).run()
# or: from websift.cli import main; main(["serve"])

Perfect for:

• VS Code (GitHub Copilot)
• Claude Desktop / Claude Code
• Cursor, Windsurf, JetBrains IDEs
• Any MCP-compatible agent

Tools

web_search(query: str) -> str

Searches DuckDuckGo and returns formatted results with title, URL, and snippet.

Example:

Agent: web_search("latest Python 3.12 features")
Server:
Title: What's New in Python 3.12
URL: https://docs.python.org/3/whatsnew/3.12.html
Snippet: Python 3.12 introduces several performance improvements...

---

Title: Python 3.12 Release Notes
URL: https://www.python.org/downloads/release/python-3120/
Snippet: The Python 3.12 release includes bug fixes and...

web_fetch(url: str) -> str

Fetches a URL and returns readable text content. Handles:

• HTML pages → Markdown (BeautifulSoup block-flow + main-content selection)
• PDF files → text extracted via pypdf only
• Plain text / JSON / XML → returned as-is
• GitHub repos → README via GitHub API (non-credential headers)
• Binary files → detected and blocked (images, executables, archives)

Example:

Agent: web_fetch("https://github.com/python/cpython")
Server:
README of https://github.com/python/cpython (via GitHub API):

# Python
The Python programming language...

Configuration

Environment Variables

Search providers

Server-wide only (SEARCH_PROVIDER). MCP tools never accept provider name, base URL, or API keys.

Convenience: pip install 'websift[providers]' (all keyed/self-hosted HTTP providers — currently no extra wheels beyond the base package; adapters use stdlib HTTP).

Optional filters (provider-dependent): SEARCH_SAFE_SEARCH, SEARCH_REGION, SEARCH_TIME_RANGE. Unsupported filters fail closed unless SEARCH_ALLOW_UNSUPPORTED_FILTERS=true.

Fallback chain (opt-in):

export SEARCH_PROVIDER=brave
export BRAVE_API_KEY=...
export SEARCH_FALLBACK_PROVIDERS=ddgs
# Does not fall back on config/auth errors (missing key, etc.)

Internal Limits

Connecting to AI Clients

1. VS Code (GitHub Copilot)

📖 Official docs: Add and manage MCP servers in VS Code | MCP configuration reference

Via Extensions View (Easiest)

1. Open Extensions view (Ctrl+Shift+X)
2. Search @mcp in the search field
3. Install any MCP server from the gallery

Via mcp.json (Custom Server)

Create .vscode/mcp.json in your workspace:

{
  "mcpServers": {
    "web-search": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Or for global (user-level) configuration, run MCP: Open User Configuration from the Command Palette and add the same entry.

Verify

Open Chat (Ctrl+Cmd+I / Ctrl+Ctrl+I) and ask: "Search for the latest Python release notes"

2. Claude Desktop

📖 Official docs: Getting Started with Local MCP Servers on Claude Desktop | Desktop Extensions

Via Settings UI

1. Open Claude Desktop → Settings → Extensions
2. Click "Advanced settings" → "Install Extension…"
3. Or manually add via the configuration file

Via Configuration File

Edit ~/.config/claude/claude_desktop_config.json (Linux) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "web-search": {
      "url": "http://localhost:8787/mcp"
    }
  }
}

Restart Claude Desktop after editing.

3. Claude Code

📖 Official docs: Connect Claude Code to tools via MCP

# Add the MCP server (HTTP transport)
claude mcp add --transport http web-search http://localhost:8787/mcp

# Verify it's connected
claude mcp list

# Use it in conversation
claude "Search for the latest Rust release and summarize the key changes"

Scopes:

# Project scope (default, stored in .mcp.json)
claude mcp add --transport http web-search http://localhost:8787/mcp

# User scope (available across all projects)
claude mcp add --transport http web-search --scope user http://localhost:8787/mcp

4. Copilot CLI

📖 Official docs: Adding MCP servers for GitHub Copilot CLI

Interactive Mode

/mcp add
# Server Name: web-search
# Server Type: HTTP
# URL: http://localhost:8787/mcp
# Press Ctrl+S to save

Command Line

copilot mcp add web-search --transport http --url http://localhost:8787/mcp

Config File

Edit ~/.github/copilot/mcp-config.json:

{
  "mcpServers": {
    "web-search": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

5. Cursor

📖 Official docs: Model Context Protocol (MCP) | Cursor Docs

Create ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "web-search": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Restart Cursor, then use Chat or Agent mode to invoke the tools.

6. Windsurf (Codeium)

📖 Official docs: Cascade MCP Integration

Create ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "web-search": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Alternatively, use the built-in UI: Settings → Cascade → MCP Servers.

7. JetBrains IDEs

📖 Official docs: MCP Server | IntelliJ IDEA Documentation

1. Install the "MCP Client" plugin from the JetBrains Marketplace
2. Go to Settings → Tools → MCP
3. Add a new server:
   • Name: web-search
   • Type: HTTP
   • URL: http://localhost:8787/mcp
4. Apply and restart

8. Any MCP Client (Generic)

For any MCP-compatible client, the server is accessible at:

• Streamable HTTP (recommended): http://localhost:8787/mcp
• SSE: http://localhost:8787/mcp/sse (set MCP_TRANSPORT=sse)
• STDIO: Run websift serve --transport stdio (or MCP_TRANSPORT=stdio websift serve)

Generic HTTP configuration:

{
  "mcpServers": {
    "web-search": {
      "type": "http",
      "url": "http://localhost:8787/mcp"
    }
  }
}

Use Cases

AI Agent Web Research

User: "Find the latest benchmarks for LLM inference optimization"
Agent: web_search("LLM inference optimization benchmarks 2025")
Agent: web_fetch("https://example.com/benchmark-article")
Agent: [Summarizes findings from fetched content]

Documentation Lookup

User: "How do I configure CORS in FastAPI?"
Agent: web_search("FastAPI CORS configuration")
Agent: web_fetch("https://fastapi.tiangolo.com/tutorial/cors/")
Agent: [Provides code example from documentation]

Error Debugging

User: "I'm getting 'ModuleNotFoundError: no module named '_sqlite3'"
Agent: web_search("ModuleNotFoundError _sqlite3 Python Docker")
Agent: [Finds solution: install python3-dev packages]

Competitive Analysis

User: "What are the latest features in React 19?"
Agent: web_search("React 19 new features")
Agent: web_fetch("https://react.dev/blog/2024")
Agent: [Summarizes new features]

Agentic AI Workflows

This server is particularly well-suited for agentic AI because:

• Deterministic tools — web_search and web_fetch have clear, predictable inputs and outputs.
• No authentication overhead — agents don't need to manage API keys.
• Self-hosted — one process/container; still needs outbound HTTPS for search and fetches.
• SSRF-safe — global-only DNS policy reduces risk of internal network exposure from agent-supplied URLs.
• Markdown output — structured text that LLMs can process efficiently.

Security

Built-in Protections

Network Considerations

• Binds to 127.0.0.1 by default. Set MCP_HOST=0.0.0.0 only when intentionally exposing (e.g. Docker); a UserWarning is emitted for non-loopback binds.
• Optional bearer auth for remote HTTP/SSE: set MCP_AUTH_MODE=bearer and MCP_BEARER_TOKEN. Clients send Authorization: Bearer <token>. STDIO does not use the token. Prefer loopback + local clients, or a reverse proxy, for production exposure.
• Search and fetch always generate outbound traffic (provider + target sites). This is not an air-gapped offline search engine.
• Docker Compose isolates the process in a container network; the image may still bind 0.0.0.0 inside the container for port publishing.

Development

Project Structure

websift/
├── pyproject.toml          # Package metadata (dynamic version), deps, console script
├── CHANGELOG.md            # Keep a Changelog
├── docker-compose.yml      # Docker Compose setup
├── Dockerfile              # Python 3.12-slim container
├── requirements.txt        # Runtime deps mirror (prefer pyproject.toml)
├── server.py               # Thin entry → websift.cli:main
├── .env.example            # Environment variable template
├── .github/workflows/      # Build/test matrix + PyPI publish gate
├── README.md               # This file
├── docs/
│   └── README.vi.md        # Vietnamese documentation
├── tests/                  # Offline pytest suite (markers: live, provider)
└── websift/
    ├── __init__.py         # WebSearchClient, AppSettings, __version__
    ├── __main__.py         # python -m websift
    ├── cli.py              # argparse CLI
    ├── settings.py         # Typed AppSettings
    ├── auth.py             # Bearer token + body limit guards
    ├── concurrency.py      # WorkLimits
    ├── models.py           # Structured internals
    ├── config.py           # Constants
    ├── security.py         # SSRF / DNS
    ├── content.py          # Content-type detection
    ├── http.py             # Page fetch
    ├── html.py             # HTML → Markdown
    ├── client.py           # Public façade
    ├── provider_http.py    # Provider credential transport
    ├── providers/          # DDGS + registry
    └── server.py           # create_server / ServerApp

Naming

Running Locally

# Install from PyPI
pip install websift

# Or editable + dev tools
pip install -e ".[dev]"

# CLI help / version
websift --help
websift --version

# Run as MCP server
websift serve

# Custom settings via CLI flags (or env)
websift serve --port 9000 --transport sse

# One-shot search / fetch
websift search "test"
websift fetch https://example.com

# Library (no server)
python -c "from websift import WebSearchClient; print(WebSearchClient().search('test'))"

Lint, test, build

# Lint
ruff check websift tests
ruff format --check websift tests

# Offline tests + coverage gate (≥85%)
python -m pytest --cov=websift --cov-report=term-missing --cov-fail-under=85 -m "not live and not provider"

# Package
python -m build
twine check dist/*

Running with Docker

# Build and start (no .env required)
docker compose up -d --build

# Optional secrets via env file or shell
# docker compose --env-file .env up -d

# View logs
docker compose logs -f

# Stop
docker compose down

Image runs as non-root (websift uid 10001), entrypoint websift, TCP healthcheck on MCP_PORT.

FAQ

Q: Does this require an API key?

Not for the default DDGS provider. DuckDuckGo search needs no key. Optional providers Brave / Tavily / Exa need server env keys (BRAVE_API_KEY, …); SearXNG needs SEARXNG_BASE_URL. Keys are never accepted via MCP tool arguments — see Search providers.

Q: Is this unlimited / no rate limits?

No. Websift does not sell a quota, but DuckDuckGo, other providers, and target sites can throttle, CAPTCHA, or block clients. Concurrent MCP calls are also bounded by SEARCH_MAX_CONCURRENCY / FETCH_MAX_CONCURRENCY / PDF_MAX_CONCURRENCY.

Q: Can I use this behind a firewall?

Yes, if outbound HTTPS to the search provider and target websites is allowed. Inbound access is only needed for the MCP endpoint when using HTTP/SSE (default loopback port 8787).

Q: How does this compare to Tavily or Firecrawl?

This is simpler and free, but doesn't offer JS rendering, deep scraping, or semantic search. For basic web search + page fetching, it's a solid free alternative. See the Comparison table for details.

Q: Can I add authentication?

Yes — for streamable-http / SSE:

export MCP_AUTH_MODE=bearer
export MCP_BEARER_TOKEN='a-long-random-secret'

Clients must send Authorization: Bearer <token>. Missing/invalid tokens get 401 without echoing the secret. STDIO ignores bearer (process-local trust). You can still put a reverse proxy in front (nginx basic auth, mTLS, etc.) and leave MCP_AUTH_MODE=none.

Optional body cap: MCP_MAX_REQUEST_BODY_BYTES=1048576.

Q: What transport protocols are supported?

• streamable-http (recommended, default) — modern MCP standard
• sse — legacy Server-Sent Events (still supported)
• stdio — for local process communication

Q: Why is the output limited to 128,000 characters?

This keeps responses within typical LLM context windows while providing substantial content. You can lower or raise it via PAGE_MAX_CHARS, WebSearchClient(max_page_chars=...), or MAX_PAGE_CHARS in websift/config.py.

Q: Can I use this for internal/private websites?

By default, SSRF protection blocks private IP ranges. To allow internal sites, modify websift/security.py to whitelist specific domains or IP ranges.

Q: Can I use this as a Python library (without MCP)?

Yes! Just pip install websift and import directly:

from websift import WebSearchClient

client = WebSearchClient(max_results=10, search_timeout=20, fetch_timeout=45)
client.search("your query")
client.fetch("https://example.com")

No server, no Docker, no MCP overhead — just pure Python.

Q: Can I publish my own version to PyPI?

Yes. After making changes:

# Install build tools
pip install build twine

# Build the package
python -m build

# Test upload (TestPyPI)
twine upload --repository testpypi dist/*

# Real upload (PyPI)
twine upload dist/*

License

MIT — see LICENSE for details.

Acknowledgments

• DuckDuckGo Search (ddgs) — search backend
• BeautifulSoup — HTML parsing
• pypdf — PDF text extraction
• FastMCP — MCP server framework
• Model Context Protocol — open protocol for AI tool integration
• VS Code MCP Documentation — official VS Code MCP guide
• Claude Code MCP Documentation — official Claude Code MCP guide
• Cursor MCP Documentation — official Cursor MCP guide
• Windsurf MCP Documentation — official Windsurf MCP guide
• Copilot CLI MCP Documentation — official GitHub Copilot CLI MCP guide
• Claude Desktop MCP Documentation — official Claude Desktop MCP guide
• JetBrains MCP Documentation — official JetBrains MCP guide