mcp-web-tools

MCP server providing web search and content extraction tools for AI assistants. Works out of the box with zero config — no API keys, no Docker. Search and extract backends are decoupled and individually swappable between SearXNG (public or self-hosted), Tavily (hosted), and Firecrawl (keyless or self-hosted).

Install

No API keys, no Docker, no config files. Just add the server — it works out of the box using a public SearXNG instance for search and Firecrawl keyless for extraction.

Claude Code (CLI):

claude mcp add mcp-web-tools -- npx -y mcp-web-tools

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "mcp-web-tools": {
      "command": "npx",
      "args": ["-y", "mcp-web-tools"]
    }
  }
}

OpenCode — add to ~/.config/opencode/opencode.json (global) or .opencode.json in your project root:

"mcp-web-tools": {
  "type": "local",
  "command": ["npx", "-y", "mcp-web-tools"]
}

That's it — search and extraction both work with zero configuration. The sections below cover customizing the backends.

Related MCP server: SSMCP

How it works

The server has two independent backends, each overridable via env vars in your client config:

• Search (WEBTOOLS_SEARCH_PROVIDER): defaults to searxng — uses a public SearXNG instance. Override with tavily, or point SEARXNG_URL at your own instance.

• Extract (WEBTOOLS_EXTRACT_PROVIDER): defaults to firecrawl keyless — no API key needed. Override with tavily, or set to none to disable web_read.

⚠️ The default public SearXNG instance is volunteer-operated. If it's down or rate-limited, set SEARXNG_URL to another instance or run your own.

Customizing backends

Use a different SearXNG instance

Point SEARXNG_URL at any SearXNG instance — public, self-hosted, or team-shared:

claude mcp add mcp-web-tools -e SEARXNG_URL=https://your-instance.example -- npx -y mcp-web-tools

Find public instances at searx.space.

Use Tavily instead (single key, both tools)

claude mcp add mcp-web-tools -e WEBTOOLS_SEARCH_PROVIDER=tavily -e TAVILY_API_KEY=your-key -- npx -y mcp-web-tools

Get a key at tavily.com. Extract automatically uses Tavily too (key reuse).

Firecrawl options

The default Firecrawl keyless tier gives ~1000 extractions/month. For higher limits:

claude mcp add mcp-web-tools -e FIRECRAWL_API_KEY=fc-your-key -- npx -y mcp-web-tools

Provider combinations

More claude mcp add one-liners:

# Point at your own SearXNG
claude mcp add mcp-web-tools -e SEARXNG_URL=http://localhost:8080 -- npx -y mcp-web-tools

# SearXNG search + Tavily extract
claude mcp add mcp-web-tools -e WEBTOOLS_EXTRACT_PROVIDER=tavily -e TAVILY_API_KEY=your-key -- npx -y mcp-web-tools

# Search only, web_read disabled
claude mcp add mcp-web-tools -e WEBTOOLS_EXTRACT_PROVIDER=none -- npx -y mcp-web-tools

# Firecrawl keyed (higher limits)
claude mcp add mcp-web-tools -e FIRECRAWL_API_KEY=fc-your-key -- npx -y mcp-web-tools

Self-hosting SearXNG (optional)

For full control over availability and rate limits, run your own SearXNG instance and point SEARXNG_URL at it.

The key requirement: SearXNG must have JSON output enabled. The default Docker image only serves HTML, so you need a settings.yml with json in search.formats. See the official docs:

→ SearXNG — Installation & Docker → SearXNG — settings.yml reference

Minimal settings.yml (mount this into the container):

use_default_settings: true
server:
  secret_key: "your-random-secret"
search:
  formats: [html, json]

Then point the server at it:

claude mcp add mcp-web-tools -e SEARXNG_URL=http://localhost:8080 -- npx -y mcp-web-tools

Full self-host (optional)

Run everything with no external network calls: SearXNG for search, Firecrawl for extract, both on your machine.

1. SearXNG — see Self-hosting SearXNG above.

2. Firecrawl — use Firecrawl's official self-host guide. Their stack needs Postgres, RabbitMQ, Redis, and Playwright:

→ firecrawl/firecrawl — SELF_HOST.md

3. Point mcp-web-tools at both:

claude mcp add mcp-web-tools -e SEARXNG_URL=http://localhost:8080 -e FIRECRAWL_URL=http://localhost:3002 -- npx -y mcp-web-tools

Tools

web_search

Search the web. Returns results with titles, URLs, and snippets. Backend is configured server-side (Tavily or SearXNG).

Parameters:

web_read

Extract clean content from web pages. Returns page text stripped of navigation, ads, and scripts. Supports up to 20 URLs per request. Backend is configured server-side (Tavily or Firecrawl). Not registered when WEBTOOLS_EXTRACT_PROVIDER=none.

Parameters:

credit_balance

Check your provider credit balance and usage. Returns "not supported" when the active provider has no credit system (e.g. SearXNG).

GitHub Tools

Three tools for reading public GitHub repository content directly via the GitHub REST API. These do not go through the search provider — they live in their own namespace and work anonymously.

github_get_repo_structure

List the directory structure of a GitHub repository path (single level, non-recursive). Useful for understanding project layout.

Parameters:

github_read_file

Read the full content of a file in a GitHub repository. Files larger than 512 KB or detected as binary are rejected.

Parameters:

github_search_repo

Search a GitHub repository across code, issues, and pull requests in one call. When kind=all (the default), all three searches run in parallel and results are grouped by type. Code search uses GitHub's /search/code endpoint, which requires a GITHUB_TOKEN (anonymous code search is not supported by GitHub — the call returns 401 without a token). Issues and PRs use /search/issues and work anonymously.

Parameters:

Environment Variables

* WEBTOOLS_EXTRACT_PROVIDER defaults to tavily when WEBTOOLS_SEARCH_PROVIDER=tavily, and to firecrawl otherwise.

Provider limitations

Each backend supports a different subset of features. Unsupported params are silently ignored.

Common pitfalls:

• SearXNG 403 — JSON output is disabled on the instance. Edit settings.yml (search.formats must include json) and restart. The docker run snippet in Backends bakes this in via the mounted settings file.

• SearXNG 429 — rate limiter is on. Disable server.limiter in settings.yml, or use a different instance.

• Firecrawl 429 — keyless free tier exhausted (~1000 credits/month, 10/min). Set FIRECRAWL_API_KEY for higher limits, or self-host.

Example Output

Node.js 22 introduces require() support for ES modules, a WebSocket client, and updates to the V8 JavaScript engine.

### 1. [Node.js — Node.js 22 is now available!](https://nodejs.org/blog/announcements/v22-release-announce)
> We're excited to announce the release of Node.js 22! Highlights include require()ing ES modules, a WebSocket client, updates of the V8 JavaScript engine, and more!

*Response time: 1.2s*

---

Sources:
- [Node.js — Node.js 22 is now available!](https://nodejs.org/blog/announcements/v22-release-announce)

Security note

This server is designed for trusted AI clients (Claude Code, Claude Desktop, OpenCode). The web_read tool accepts any http(s):// URL supplied by the calling client and does not implement SSRF protections (e.g. blocking of private/internal IP ranges). Do not expose this server to untrusted input.

License

MIT