How to Connect Perplexity MCP to Claude (2026 Setup Guide)

Claude’s training data has a cutoff, and even Claude’s built-in web search is a single-turn tool — you ask, it searches once, done. If your workflow involves ongoing research-heavy work — market research, competitive analysis, fact-checking long drafts — connecting Perplexity’s official MCP server gives Claude a dedicated, persistent search layer powered by Perplexity’s Sonar models, with citations attached automatically to every result.

This guide covers the full setup, what it actually changes versus Claude’s native search, common errors, and whether it’s worth the extra step for you.

What Perplexity MCP Actually Does

The Model Context Protocol (MCP) is an open standard that lets Claude call external tools directly instead of relying only on what it already knows. Perplexity’s official MCP server exposes its Sonar search and reasoning models — including sonar-pro and sonar-deep-research — as tools Claude can call mid-conversation.

Practically, this means Claude can run a live, cited web search whenever it needs current information, and can chain multiple searches together for multi-step research tasks — something a single search call can’t do.

Perplexity MCP vs. Claude’s Built-In Web Search

Worth clarifying before you set anything up, since they overlap:

• Claude’s native web search is free, built-in, and requires zero setup — fine for quick fact-checks or “what’s the latest on X.”
• Perplexity MCP requires a paid API key (billed separately from a Perplexity Pro subscription) but gives you Perplexity’s specific Sonar models, including a deep-research mode built for exhaustive multi-source investigation, and lets you keep a consistent research tool across other MCP clients (Cursor, VS Code, Claude Code) beyond just Claude.

If you’re a casual user, native search covers you. If you’re doing structured research work regularly and already pay for Perplexity’s API, the MCP server is worth the five minutes.

What You’ll Need

• A Perplexity API key (from the API Portal — this is separate from a Perplexity Pro/Max subscription; API access is billed per use)
• Node.js installed (the server runs via npx)
• Claude Desktop, Claude Code, or another MCP-compatible client

Step 1: Generate Your API Key

Log into Perplexity’s API Portal and generate a new key. Copy it somewhere safe — you’ll paste it into your config in the next step, and it won’t be shown again after you navigate away.

Step 2: Install the Server

Fastest option, via Claude Code’s CLI:

claude mcp add perplexity --env PERPLEXITY_API_KEY="your_key_here" -- npx -y @perplexity-ai/mcp-server

For Claude Desktop, add this block to your config file instead:

{
  "mcpServers": {
    "perplexity": {
      "command": "npx",
      "args": ["-y", "@perplexity-ai/mcp-server"],
      "env": {
        "PERPLEXITY_API_KEY": "your_key_here"
      }
    }
  }
}

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Step 3: Restart Claude

Fully close and reopen Claude Desktop, or start a new Claude Code session, so it picks up the new server.

Step 4: Verify the Connection

In Claude Code, list your MCP servers and confirm Perplexity shows as connected. In Claude Desktop, just ask something time-sensitive — a working connection returns a current, cited answer instead of a training-data answer with no sources.

Common Setup Errors

• “Tool not found” after install: usually means the package didn’t install correctly — re-run the npx command manually in a terminal to see the actual error before troubleshooting further.
• Timeout on deep research queries: sonar-deep-research can genuinely take 30–60+ seconds for multi-step investigation — this isn’t a broken connection, just a slow model by design.
• Initialize errors on strict clients: some MCP clients fail if npx prints install logs to standard output; using npx -yq instead of npx -y suppresses that noise.
• Nothing happens after restart: double-check the JSON in your config file is valid — a missing comma or bracket silently breaks the whole mcpServers block, not just the Perplexity entry.

Where This Actually Helps

• Fact-checking a long draft against current sources before publishing, instead of manually cross-checking claim by claim
• Competitive research where you want Claude to pull pricing or feature updates from multiple live sources in one pass
• Multi-step research projects where Claude needs to search, read, then search again based on what it found — something single-shot search can’t do

Is It Worth Setting Up?

If you already pay for Perplexity API access or do frequent research-driven work inside Claude, yes — it’s a genuinely different, more persistent search layer than what’s built in. If your searches are occasional and casual, Claude’s native web search will cover you without the extra API key to manage.