Most search tools give you a query box and one behaviour; here the model, the reasoning effort and how much context is read are arguments, so the same server serves a five-second fact check and a long comparative piece without reconfiguration. Since all three settings move both latency and cost, the useful habit is the one the tool itself recommends: gpt-5-mini at low effort for iterating, and the larger model with higher effort reserved for the questions that earn it.
A single-tool server over OpenAI's web search. One call carries the query plus the knobs that decide how hard the model works on it: which model, how much reasoning effort, how much search context, and where the searcher is.
- Model chosen per call across the reasoning line — gpt-5, gpt-5-mini, gpt-5-nano, o3, o4-mini — as well as gpt-4o and gpt-4o-mini, defaulting to gpt-5-mini
- reasoning_effort of minimal, low, medium or high, so a quick check and a deep research pass are the same tool with a different setting
- search_context_size of low, medium or high, defaulting to medium — how much the model reads before answering
- user_location for results that depend on where you are
- The tool's own guidance built in: gpt-5-mini with low effort for fast multi-round searching, gpt-5 with medium or high for deep research
- OPENAI_DEFAULT_MODEL sets the default so routine calls do not have to name one
An OpenAI API key from the OpenAI platform, in OPENAI_API_KEY — searches are billed against it, and effort and context size are what move that cost. It is a Python package run with uvx openai-websearch-mcp, so uv is the only local dependency; uvx --with openai-websearch-mcp openai-websearch-mcp-install writes the Claude Desktop configuration for you. MIT.
One command plus a key — OPENAI_API_KEY=sk-xxxx uvx --with openai-websearch-mcp openai-websearch-mcp-install, then supply credentials