What you need before you start
The question “can you connect Perplexity to Claude” covers two opposite setups, and they share only the protocol.
The first gives Claude a Perplexity search tool: Claude calls Perplexity when it wants fresh web results, a reasoned answer or a research pass. That is what Perplexity’s own MCP server does, and it is the subject of this guide.
The second turns Perplexity into the client: you attach other MCP servers to Perplexity so it can reach your tools while it answers. That is a different configuration, documented in Perplexity’s help centre, and it does not involve Claude at all.
For the first setup, three things are needed:
- A Perplexity API key, from the API console at console.perplexity.ai. This is not the login of a Pro subscription: a Pro plan and an API key are separate products, and the key is billed per use against a credit balance.
- An MCP client, Claude Desktop, Claude Code, Cursor, VS Code, Windsurf or Kiro, since the server ships install buttons for several of them.
- Node.js, if you take the local route rather than the remote one.
That billing point is the one that surprises people. A Pro subscription at $17 a month billed yearly buys the Perplexity app, not API calls, so the MCP server is a paid-by-use tool rather than something the subscription already covers.
How the job is done with Perplexity
Perplexity hosts a remote server, and it is the shortest path: nothing to install, nothing to update.
The remote server. The address is https://api.perplexity.ai/mcp, over Streamable HTTP, with your API key in the authorization header. In Claude Code:
claude mcp add --transport http perplexity https://api.perplexity.ai/mcp --header "Authorization: Bearer YOUR_KEY"
The local server. The same tools, run on your machine through npm, which suits a client that cannot hold a remote server or a network that blocks one. In Claude Code:
claude mcp add perplexity --env PERPLEXITY_API_KEY="YOUR_KEY" -- npx -y @perplexity-ai/mcp-server
Claude Desktop and other clients. Add this block to the client configuration file, the name of which depends on the client: claude_desktop_config.json for Claude Desktop, ~/.cursor/mcp.json for Cursor, ~/.codeium/windsurf/mcp_config.json for Windsurf, .vscode/mcp.json for VS Code.
{"mcpServers": {"perplexity": {"command": "npx", "args": ["-y", "@perplexity-ai/mcp-server"], "env": {"PERPLEXITY_API_KEY": "YOUR_KEY"}}}}
The plugin route. In a Claude Code session, the repository is packaged as an agent plugin: add the marketplace with /plugin marketplace add perplexityai/modelcontextprotocol, then install it with /plugin install perplexity. Secrets do not travel through the plugin, so the key is still set in the environment.
What the server exposes is three capabilities rather than three endpoints: real-time web search from the Search API, a reasoning pass, and a research mode that runs through the Agent API. In practice, Claude stops guessing about recent events and asks Perplexity instead.
On this page
What to check in the result
Ask a question whose answer changed in the last week, and check that the reply carries sources with dates rather than a confident summary. A Perplexity tool call that returns no citations is not reaching the search API.
Watch the timeout on a research query. The local server waits five minutes by default, and the environment variable PERPLEXITY_TIMEOUT_MS raises it, which matters on long research runs.
Check which key is in play before blaming the model. A revoked key produces an authentication error at connect time, while a key with no credit produces it at the first call, and the two look similar in a chat window.
Where it usually goes wrong
An API key taken from the wrong place. The Perplexity subscription page does not issue API keys; the console does, and a login session is not a key.
A Pro subscription mistaken for API credit. The connection authenticates, the first search fails, and the reason is a balance rather than a configuration.
Node missing on the machine. The local route fails at launch with a command-not-found error before Claude sees anything.
A client that cannot hold remote servers. The remote address is the documented default, and the local server is the fallback rather than the other way round.
A corporate proxy. The local server reads PERPLEXITY_PROXY first, then the standard HTTPS_PROXY and HTTP_PROXY, and a proxy that is configured in the browser but not in the environment leaves the server with no route out.
What the plan allows
On the Perplexity side, the MCP server is an API product. Pro at $17 a month, billed yearly, opens the app, more Computer usage, deep research and higher limits. Max at $167 a month, also billed yearly, goes further on the same lines. Neither figure covers API calls, which are billed separately against the key.
On the Claude side, the client decides. Claude Code needs a Pro or Max subscription. Claude Desktop and the other clients do not, since the server is a local process or a remote address they hold like any other.
So the real cost of this connection is the API spend, and it follows the number of searches and the depth of each research run rather than a monthly tier.
The full grid, tier by tier and with the date each price was read, is in the Perplexity review. ## What we did not do
Nobody on this site has connected Perplexity to Claude. We did not create an API key, add the server to a client, run a search through it or measure what a research query costs. Everything above comes from Perplexity’s MCP repository, read on November 16, 2026, and from the pricing page read on October 30, 2026.
The steps are documented rather than tested. We did not verify the tool names as they appear in a Claude session, we did not measure latency for either route, and we did not test the plugin installation path.
One limit to state plainly: the API is priced per use, and this guide gives no figure per search. The cost of a session depends on how many calls Claude decides to make, which is a property of the conversation rather than of the price list.
Sources
Read on November 16, 2026.
Perplexity API Platform MCP Server, https://github.com/perplexityai/modelcontextprotocol.
MCP integration, https://docs.perplexity.ai/docs/getting-started/integrations/mcp-server.
API console, https://console.perplexity.ai.
Pricing, https://www.perplexity.ai/pro, read October 30, 2026.
Update log
November 16, 2026: first version, built from Perplexity’s MCP repository and pricing page. No hands-on test has been done.
Start with Perplexity
This is an link to the publisher. A commission may be paid if you subscribe, and the price stays the same for you.