SearchForge

What divyanshu-iitian/SearchForge MCP server does

The divyanshu-iitian/SearchForge MCP server gives MCP-compatible agents a local retrieval gateway for search and web reading. It can route queries through GitHub repository search, Crossref metadata, Hacker News through Algolia, Wikipedia, and an optional self-hosted SearXNG deployment. A URL reader converts public web pages into Markdown through Jina Reader.

The server does not generate answers or conceal citations. Its role is retrieval: return search results with source information, fetch readable page content, and report the current condition of configured providers. No telemetry is sent according to the project documentation.

How it works

Search uses an auto category by default. That routing mode identifies code, academic, and current or community-oriented signals, then queries relevant source families with a web fallback. Explicit categories can narrow the provider set, while an explicit provider list overrides category routing for evaluation workflows.

Results are deduplicated and rankings are combined across providers. Search responses include structured result data and provider diagnostics such as latency, success state, and result counts. The MCP tools are exposed over stdio, and the same retrieval layer is also available through a CLI, REST API, and TypeScript package.

The read operation accepts public HTTP and HTTPS URLs only. Requests to localhost, private IP literals, credential-bearing URLs, and non-web protocols are rejected. Responses are bounded in size, subject to timeouts, and cached.

Setup and configuration

The project requires Node.js 20 or newer. A local setup can be built by cloning the repository, installing dependencies, and running the build command. The MCP process can then be started with the generated dist/mcp.js file. A client may also launch it through npx directly from GitHub, or use the published OCI image with a client that supports Docker commands.

Set SEARCHFORGE_SEARXNG_URL when the server should use a private SearXNG instance. The bundled Docker Compose setup starts SearchForge on port 3000 and SearXNG on port 8080. Before exposing that stack, the documentation recommends changing the SearXNG secret, setting SEARCHFORGE_API_KEY, and placing TLS termination behind a trusted proxy.

Tools and capabilities

The divyanshu-iitian/SearchForge MCP server exposes three stdio tools:

• web_search performs routed, structured searches with source details.
• read_url retrieves a public page as clean Markdown.
• search_status runs live capability and latency checks for the retrieval path.

Search categories include automatic routing, web, code, academic, and community use cases. GitHub and Crossref work without keys in the documented default setup; a GitHub token can increase the public API quota. Brave is an optional keyed backend, while SearXNG is self-hosted.

Limitations and notes

The read tool is limited to public HTTP(S) URLs and the Jina Reader path is currently rate-limited. Provider failures are reported as degraded capability rather than an all-or-nothing health result. The default web capability includes Wikipedia, while broad independent web metasearch requires the private SearXNG stack. The project is licensed under MIT.