Perplexity MCP Server

What is Perplexity MCP Server?

An MCP server that provides Perplexity AI web search capabilities to Claude, with automatic model selection, stateful filters, and 10 purpose-built tools. It is designed for users of Claude Desktop or any MCP-compatible client who need AI-powered search, deep research, raw web results, agentic loops, and embeddings.

How to use Perplexity MCP Server?

Install Node.js v20 or higher, obtain a Perplexity API key, and ensure you have Claude Desktop (or any MCP-compatible client). Clone the repository, run npm install and npm run build. Add the server to Claude’s config file (claude_desktop_config.json) with the command, absolute path to the built index.js, and environment variables PERPLEXITY_API_KEY and PERPLEXITY_MODEL. After restarting Claude, you can invoke the search, raw_search, async_research, agent, embeddings, and filter tools.

Key features of Perplexity MCP Server

• 10 purpose-built tools covering search, raw results, research, agent, and embeddings
• Automatic model selection based on query keywords (sonar, sonar-pro, sonar-reasoning-pro, sonar-deep-research)
• Stateful domain allowlist/blocklist and recency time-window filters
• Asynchronous deep research jobs that can be polled, surviving 7 days
• Agentic loop with built-in tools (web_search, fetch_url) and optional third-party models
• Embeddings generation via pplx-embed models with Matryoshka dimensions

Use cases of Perplexity MCP Server

• Time-sensitive research with domain filtering: set recency to week, allow only specific domains, then search for recent breakthroughs
• Financial document research: use raw_search with SEC mode to find filings, then search for synthesized analysis
• Academic literature review: search with academic mode and high context size for peer-reviewed sources
• Deep research with reasoning control: run search with high reasoning effort and strip thinking blocks for clean output
• Quick fact-checking: use raw_search to get ranked results without AI synthesis, cheaper and faster

FAQ from Perplexity MCP Server

What are the prerequisites for Perplexity MCP Server?

Node.js v20 or higher, a Perplexity API key (get one at https://www.perplexity.ai/settings/api), and Claude Desktop or any MCP-compatible client.

How does automatic model selection work?

The server scores your query against keyword lists. Research keywords (e.g., "deep research", "comprehensive") trigger sonar-deep-research; reasoning keywords (e.g., "solve", "logic") trigger sonar-reasoning-pro; simple keywords (e.g., "quick", "brief") trigger sonar; everything else uses sonar-pro. A strong match (score ≥ 2) overrides a manually set model.

What models are available and how can I set a default?

Available models: sonar-deep-research, sonar-reasoning-pro, sonar-pro, sonar. Set a default via the PERPLEXITY_MODEL environment variable, or use the model_info tool to view and override the current selection.

Can I use streaming responses?

Yes, the search tool accepts a stream boolean parameter and a stream_mode parameter (full or concise) for Pro Search streaming events.

How do filters work and are they persistent?

Domain filters (allowlist/blocklist) and recency filters persist across all subsequent searches until cleared. Use clear_filters to reset all, list_filters to view active filters. Maximum 20 domains; cannot mix allow and block in the same filter set.