Perplexity MCP: search, ask and run deep research from an agent

Some questions take more than one search, like “compare how three vector databases changed their pricing over the past year”. That means rounds of searching, reading and summarizing, and a coding agent doing it itself is slow and burns context. Perplexity MCP hands that whole stretch to Perplexity: the agent asks, and gets back a sourced answer or research report. It’s maintained by Perplexity and had about 2.6k stars as of 2026-10-06.

What it does

• Raw search: perplexity_search calls the Search API and returns ranked results with metadata, with recency and domain filters, plus a search_type: "fast" option that trades some depth for lower latency and cost.
• Quick answers: perplexity_ask answers everyday questions with live search, backed by the Agent API’s fast preset.
• Reasoning: perplexity_reason handles analytical problems with the medium preset.
• Deep research: perplexity_research uses the high preset; a run can take minutes, and the server reports progress to clients that ask for it. All four tools sit on managed presets, a model, retrieval setup and step budget that Perplexity keeps tuning, rather than a hard-coded model name.

Who it’s for

• People writing tech evaluations or market analyses who want to offload multi-round searching so the main agent can focus on writing.
• Developers already on the Perplexity API who want it inside Claude Code or Cursor.

Setup

Needs a Perplexity API key from the API console. The hosted server is the easiest path; Claude Code takes one command:

claude mcp add --transport http perplexity https://api.perplexity.ai/mcp --header "Authorization: Bearer YOUR_API_KEY"

To avoid depending on the hosted endpoint, run it locally instead with npx, Docker, or a plain Node process in HTTP mode on your own machine or server. The local server also supports a configurable timeout, a custom base URL, and PERPLEXITY_PROXY for corporate proxy setups.

Our take

Most search servers return results and leave reading and synthesis to your agent; Perplexity is one of the few official servers that outsources the research itself, with four tools graded from light to heavy so you can match the question — unlike Exa MCP, whose default tools hand back cleaned page content rather than a synthesized report. Two catches: every call is billed by API usage, and perplexity_research costs noticeably more per run; and the answers are Perplexity’s synthesis, so verify key claims through the cited sources. Perplexity tunes the presets over time, so the same question may get different results later, and the legacy model names (sonar-pro and similar) and parameters like reasoning_effort that older versions of this server exposed are gone from the current tool schemas and are silently ignored if sent. Licensed MIT.