What it takes off you is guessing how much to spend on a question. The fast preset handles the factual ones, the research preset is there when a question deserves the wait, and the plain search call exists for when a summary would be in the way.
Four tools over the Perplexity Agent API, separated by effort rather than by subject: `fast` for an answer, `medium` for one that reasons, `high` for a research pass, and plain ranked search underneath them.
- A web-grounded conversational answer on the `fast` preset, for quick questions and everyday searches (`perplexity_ask`)
- Advanced reasoning on the `medium` preset, for comparisons, calculations and analytical work (`perplexity_reason`)
- Deep research on the `high` preset for thorough analysis and detailed reports — runs can take minutes, and the server streams progress to clients that ask for it (`perplexity_research`)
- Ranked search results with metadata, returned without AI synthesis, for when you would rather read the sources yourself (`perplexity_search`)
- Recency and domain filters on the search call (`search_recency_filter`, `search_domain_filter`)
- Presets managed by Perplexity — model, search setup and step budget — rather than a model name you pin and have to revisit
A Perplexity API key from the API console. The easiest path is Perplexity's own remote server at `https://api.perplexity.ai/mcp` over Streamable HTTP with the key in an `Authorization: Bearer` header — same tools, nothing to install or update. Locally it is `npx -y @perplexity-ai/mcp-server` with `PERPLEXITY_API_KEY` set, and `PERPLEXITY_TIMEOUT_MS` raised from its five-minute default if you intend to use the research preset. One migration note: earlier versions called the `sonar-pro`, `sonar-reasoning-pro` and `sonar-deep-research` models and accepted `strip_thinking` and `reasoning_effort`; those parameters are gone from the schemas and are ignored if sent.
One command plus a key — npx -y @perplexity-ai/mcp-server, then supply credentials