Web Search

What web-search does

The web-search MCP server exposes a single search tool for querying six search engines: Serper, Brave, Exa, Tavily, Firecrawl, and Perplexity. Agents can use it for general web research or select a mode for web, news, academic, answer, or page-oriented requests.

Each response contains ranked results with URLs, titles, snippets, and optional content. Results can also include publication times and the engine that supplied them. In answer mode, the response may include a synthesized answer with citations. Metadata reports a request ID, cache tier, whether the request was degraded, and the billed cost.

How it works

The hosted service provides one streamable-HTTP MCP endpoint at https://api.groundroute.ai/mcp. GroundRoute chooses an engine using price and a quality threshold, then caches responses for repeat queries. If an engine becomes unavailable or degrades, the service can route the request to another engine without requiring changes in the agent integration.

The web-search MCP server supports bring-your-own-key usage, while the hosted service handles the provider routing layer. The repository also includes a benchmark covering 170 agent queries across the six engines; the README links to the methodology and results.

Setup and configuration

Create an API key at groundroute.ai/keys, then connect an MCP client to the hosted endpoint with an Authorization: Bearer header. The README provides configuration examples for Claude Desktop, Claude Code, Cursor, VS Code, and Continue.

Clients that only support stdio can use mcp-remote to bridge the hosted HTTP endpoint. Alternatively, clone or download the repository and run its small native stdio wrapper with Python 3.12 or later:

pip install -r requirements.txt
GROUNDROUTE_API_KEY=gr_YOUR_KEY python server.py

The wrapper can also run in Docker. Tool discovery works without a key, but executing a search requires GROUNDROUTE_API_KEY.

Tools and capabilities

The server exposes one tool, search. Its required argument is query. Optional arguments include:

• mode: auto, web, news, academic, answer, or page
• max_results: defaults to 10 and accepts up to 50
• freshness: fresh, semi, or static
• domains: an include-only list of domains
• lang: an ISO 639-1 language code
• country: an ISO 3166-1 alpha-2 country code

Use domain filtering for focused sources such as arxiv.org, and set language or country when regional search results matter. The web-search MCP server is a fit for agents that need structured search output, optional synthesized answers, and visibility into the selected engine, cache tier, and request cost.