Perplexity MCP Server Setup for Cursor IDE (2026)

How do you set up the Perplexity MCP server in Cursor? Generate an API key from Perplexity's developer settings, add a perplexity-ask block to ~/.cursor/mcp.json with the key in the env section, and restart Cursor. The model can then ask Perplexity a question and get back a synthesized, cited answer built on live web search — not a list of links to click through and read yourself.

That's the real difference from a generic search server: this one is grounded in Perplexity's own search-and-synthesis models, so what comes back reads like an answer with sources attached, rather than raw result snippets the AI still has to interpret.

Perplexity vs. a Plain Search Server

This site also covers the Brave Search MCP server, and the two solve overlapping but genuinely different problems. Brave Search returns ranked results — titles, URLs, snippets — and it's Cursor's own model doing the reading and summarizing afterward. Perplexity does that synthesis step itself before the answer ever reaches Cursor: you get a direct answer with citations, already reasoned over multiple sources.

The trade-off runs the other way, too. A raw result list from Brave lets the model pick which sources to trust and cross-reference them itself. Perplexity's synthesized answer is faster to read but hands you someone else's synthesis — worth a second check with a plain search when the stakes are high enough that you want to see the underlying sources yourself rather than trust a pre-built summary.

Get a Perplexity API Key

1. Go to Perplexity's API settings in your account dashboard (perplexity.ai, under API access)
2. Generate a new API key
3. Perplexity's API is metered/paid usage rather than a free consumer-plan feature — check current pricing before wiring this into a workflow you'll run often, since per-query costs vary by which underlying model handles the request
4. Copy the key somewhere safe

Add It to Cursor

Open ~/.cursor/mcp.json (or Settings → MCP in Cursor) and add:

{
  "mcpServers": {
    "perplexity-ask": {
      "command": "npx",
      "args": ["-y", "server-perplexity-ask"],
      "env": {
        "PERPLEXITY_API_KEY": "your-perplexity-api-key-here"
      }
    }
  }
}

Restart Cursor completely for the new server to register.

Confirm It's Working

In Cursor chat:

Ask Perplexity what changed in the most recent stable release of Node.js, with sources

A direct answer citing specific sources — not a raw list of search results — confirms the connection works and it's genuinely hitting Perplexity's synthesis, not just doing a plain web search.

Practical Workflows

Getting a straight answer instead of a link dump

What's the current recommended way to handle authentication in Next.js App Router — ask Perplexity and give me the direct answer with sources, not just links

Fact-checking something before it goes in documentation

I want to write that this library dropped support for Node 16. Ask Perplexity to confirm that's actually accurate and cite where it's documented

Catching up on something that moved fast

Ask Perplexity to summarize what's changed in the Postgres 17 release notes that would actually affect a typical web app's database layer

Comparing current options with sources attached

Ask Perplexity for a sourced comparison of the current hosted Postgres options — what's genuinely different between them right now, not marketing copy

What This Doesn't Do

It's not a general-purpose research browser — you get a synthesized answer, not the ability to have the AI read one specific long document start to finish. If you need Cursor to ingest a specific URL's full content rather than get an answer about a topic, that's a fetch-style server's job, not this one. It also isn't free the way some reference MCP servers are; every call is a metered API request against your Perplexity account, so it's worth being deliberate about when a question actually needs sourced synthesis versus when the model's own training-data knowledge is good enough.

Troubleshooting

"Invalid API key" on every request
Double-check the key was pasted into mcp.json without extra whitespace or a trailing newline, and confirm it's still active in your Perplexity dashboard — keys can be individually revoked without closing the account.

Answers come back but without visible citations
Ask explicitly for sources in the prompt itself — "with sources" or "cite where this comes from" — rather than assuming citations are always surfaced by default in how Cursor renders the response.

Requests are slower than a plain search server
That's expected — synthesis takes longer than returning raw snippets, since the underlying model is reading multiple sources and composing an answer before responding. If speed matters more than a synthesized answer for a given task, a plain search server will feel faster.

Getting confidently wrong answers on a fast-moving topic
Synthesis is only as good as what's indexed and how current it is. For anything published in the last few hours, spot-check the citation yourself — very recent information is where any search-based tool is most likely to be thin or slightly behind.

Costs adding up faster than expected
Each query is a metered API call. If a workflow chains several Perplexity calls per session without you noticing, check your usage dashboard — it's easy to burn through a budget faster than a single "ask a question" mental model suggests.

Frequently Asked Questions

Q: Is Perplexity MCP the same as using Brave Search or Exa?
A: No. Brave Search and Exa return search results for the model to read and summarize itself. Perplexity does that synthesis step on its own end and hands back a direct, cited answer — different output shape, different trade-offs around speed and transparency of sources.

Q: Is the Perplexity API free?
A: No, it's metered/paid usage tied to your account, unlike some reference MCP servers that only cost you the underlying free-tier API. Check current pricing before building a workflow around frequent calls.

Q: Can I control which underlying model Perplexity uses for a query?
A: Perplexity's API exposes different models with different cost and quality trade-offs; check the current API documentation for exact model names and how to specify one, since this is an area providers update relatively often.

Q: Does this replace needing the model's own reasoning?
A: No. It supplies grounded, current information — the model still has to decide when a question needs that grounding versus when its own knowledge is sufficient, and still does whatever reasoning or code-writing happens after the answer comes back.

Q: Why would I use this instead of just asking Cursor's model directly?
A: The model's own training data has a cutoff and can be wrong about anything that changed since. Perplexity queries the live web and cites sources, which matters for anything time-sensitive — current versions, pricing, recent incidents, anything published after the model's training cutoff.

Related Guides

---