Web Retrieval MCP

What web-retrieval-mcp MCP server does

The web-retrieval-mcp MCP server exposes read-only web research tools over Model Context Protocol. Its documented core operations are web_search and web_fetch. Search results are returned as separate blocks containing fields such as the title, URL, publication date, highlights, and text, followed by a Sources list. Exa deep search modes can also return a labeled synthesized answer and grounding when the provider supplies it.

Page retrieval supports a full readable body, a shorter summary, or an answer grounded in a question about the page. Responses include provenance information, including which retrieval tier served the content. The package also includes tools for discovering AI and machine-learning papers and searching developer material such as documentation, READMEs, issues, and merged pull requests.

How it works

Search can use Exa or Tavily, with Firecrawl available as a disclosed fallback. Full-page retrieval follows a tiered route: indexed content is attempted first, then a local Camoufox browser can handle rendered pages, followed by optional Tavily Extract and Firecrawl paths. The selected route depends on the request and the configured providers and optional packages.

The web-retrieval-mcp MCP server accepts controls for freshness, date ranges, domains, categories, result counts, rendering, and per-result text size. Fetch responses identify the serving tier, cache state, semantic mode, and whether content was truncated. Unsupported provider selections, missing optional dependencies, and unavailable credentials are reported in tool results rather than silently ignored.

The server can run locally over stdio or as a loopback HTTP service. An optional Valkey sidecar can cache completed fetch results, and the server validates initial URLs and guards browser requests and redirects against SSRF risks.

Setup and configuration

Python 3.10 or newer is required. A minimal installation is available with pipx install web-retrieval-mcp; uvx web-retrieval-mcp can run it in a temporary isolated environment. The all extra adds the optional Tavily, local rendering, Valkey cache, and native key-store capabilities. Installing the rendering extra and using Camoufox requires running camoufox fetch to download its managed browser.

Exa and Firecrawl credentials provide the broadest routing coverage. Tavily is optional and adds provider-selectable search and another extraction tier. The documented environment variables are EXA_API_KEY, FIRECRAWL_API_KEY, and TAVILY_API_KEY. Credentials are resolved lazily, so a missing key is reported when a selected route needs it. The README also describes key-file and operating-system secret-store alternatives.

Claude Desktop, Cursor, Codex, Claude Code, and other clients accepting MCP server configuration can launch the executable directly. Clients without it on PATH can use uvx with web-retrieval-mcp as the argument.

Limitations and notes

Provider availability and capabilities depend on installed extras and configured credentials. Local rendered retrieval requires the Camoufox package and browser download. Tavily mappings, Firecrawl fallback filter gaps, and some Exa category or filter changes can affect requests and are surfaced to callers or server diagnostics.

The server is intended as a separate retrieval lane and can complement built-in agent web tools. The package includes an optional previewable hook installer for replacing Claude Code's built-in web tools, but that behavior is not required for normal MCP use. Provider pricing and entry allocations can change, so current provider terms should be checked before estimating usage.