Passing what you are trying to find out, alongside the search terms, lets the server select excerpts against the goal rather than against keyword overlap. That is the difference between a snippet that happens to contain the word and one that answers the question. Splitting fast search from authenticated deep research is the other sensible division: the cheap path stays open and anonymous.
Parallel publishes two endpoints. The search one answered our probe unauthenticated with two tools; a separate Task endpoint handles deep research and enrichment and always needs credentials.
- web_search — low-latency search returning LLM-friendly results with excerpts, guided by a stated objective
- web_fetch — extract content from specific URLs, documented as a follow-up for when search excerpts are not enough
- On the Task endpoint: createDeepResearch, createTaskGroup, getStatus, getResultMarkdown
Nothing for search. The Task endpoint needs an API key or OAuth. One-click install links exist for Cursor, VS Code, LM Studio and Goose.
One command — npx -y mcp-remote https://search.parallel.ai/mcp