What it is
Perplexity hosts the server at https://api.perplexity.ai/mcp over Streamable HTTP, authenticated with an API key sent as a bearer token. It can also run locally with npx -y @perplexity-ai/mcp-server (version 1.3.0, published in September 2026) and a PERPLEXITY_API_KEY, be self-hosted in HTTP mode with Docker, or be embedded as a library in your own Node process.
It has four tools. perplexity_search calls the Search API and returns ranked results with dates, recency and domain filters. perplexity_ask, perplexity_reason and perplexity_research call the Agent API at its fast, medium and high effort levels and return answers with citations; research runs can take minutes and stream progress. Earlier versions called the Sonar models, and their strip_thinking and reasoning_effort parameters are now ignored.
The repository is MIT-licensed, has about 2,600 stars and is listed in the official MCP Registry as ai.perplexity/mcp-server.
Install
https://api.perplexity.ai/mcpOur review
Most search servers hand your model links and snippets to read. This one can too, but its distinctive tools return finished answers with citations, so the main agent receives a few hundred words instead of ten pages of results. That saves context in the calling model and moves the reading cost to Perplexity's bill.
Which tool the agent picks decides the cost. perplexity_search is flat: $5 per 1,000 requests, or $1 on Fast Search, and failed or rate-limited requests are not billed. The answer tools bill the underlying model's tokens plus each tool call, and Perplexity's effort-level docs allow up to 15 steps at medium and high, each of which can include a search or a URL fetch. Perplexity publishes no per-request figure; responses report a total_cost field after the fact.
Rate limits matter more than the README suggests. Agent API limits rise with lifetime spend: a new account gets 1 request per second for the whole organisation, $50 of purchased credit lifts that to 3, and $250 to 8. The Search API allows 50 query units a second on every tier. A shared team key on a fresh account will queue under modest agent load.
On data, Perplexity says API content is not used to train models, and holds a SOC 2 Type II report. Zero data retention is off by default and offered only under a written agreement, and the presets behind the answer tools run third-party models (OpenAI's, at the time of writing) whose providers may require limited retention. Because Perplexity retunes the presets over time and the MCP tools expose no version pin, the model behind perplexity_research today is not guaranteed to be the one behind it next quarter.
Where it earns its place
A solo developer working in Claude Code
Check current library versions and breaking changes before an upgrade
perplexity_search with a recency filter returns dated results the model can cite; at $5 per 1,000 requests, 2,000 lookups a month costs $10, or $2 on Fast Search. Results are untrusted web content, so the agent should not act on instructions inside them.
An analyst at a 10-person market-research firm using Claude Desktop
Produce a cited first-draft brief on a competitor before a client call
perplexity_research runs the high effort level for several minutes and returns a sourced report; the bill is tokens plus every search and fetch it makes. Citations still need checking, and zero data retention requires a signed agreement.
A platform team running shared MCP servers for 200 internal users
Offer web search to every internal agent without each person managing a key
Self-host the HTTP mode, or embed the library with a per-tenant key resolver, setting BIND_ADDRESS and ALLOWED_HOSTS deliberately. Buy at least $250 of credit up front to reach tier 2 and 8 Agent API requests a second.
Before you commit
- Setup
- One command for the hosted server: claude mcp add --transport http perplexity https://api.perplexity.ai/mcp with a bearer key; npx for local use, plus a Claude Code plugin and one-click installs for Cursor and VS Code. Configs are documented for Claude Desktop, Codex, Kiro and Windsurf.
- Integrations
- Four tools, all read-only: perplexity_search (ranked results, recency, domain and fast-search options), perplexity_ask (fast), perplexity_reason (medium) and perplexity_research (high). ask and reason accept search-context budgets up to 120,000 tokens.
- Data and privacy
- API content is not used for training; zero data retention only by written agreement; SOC 2 Type II. The hosted server sees every query, and the tools fetch untrusted web pages, so prompt injection through results is possible; an open issue proposes flagging untrusted sources.
- Cost at scale
- 10,000 raw searches a month cost $50, or $10 on Fast Search. Answer tools add model tokens plus $0.0025 per standard web search, $0.001 per fast search and $0.0005 per URL fetch, so a research call that makes a dozen searches costs about three cents before tokens.
- Limits
- Agent API: 1 to 33 requests a second by spend tier; Search API: 50 a second. The local server times out after 5 minutes by default, which long research runs can exceed, and some strict clients need npx -yq to keep install messages off stdout.
What's good
- Cited answers rather than raw links, saving the caller's context
- Search API failures and 429s are not billed
- Hosted, local, self-hosted and library options
- Self-hosted HTTP mode binds to localhost with CORS closed by default
What's not
- Answer-tool cost per call not published in advance
- New accounts limited to 1 Agent API request per second
- No free anonymous tier; needs a key with credit
- Presets change underneath you with no pin in the MCP tools
Best for
Developers and analysts who want an assistant to return cited answers and research briefs from the live web.
Alternatives
Perplexity MCP Server: common questions
- Is Perplexity MCP Server worth it?
- Worth its premium when you want a second model to do the reading and hand back citations; for raw links, plainer search servers cost less. We rate it 4.2 out of 5.
- How much does Perplexity MCP Server cost?
- Server free. Search API $5 per 1,000 requests ($1 with Fast Search); answer tools bill Agent API model tokens plus $0.0025 per web search and $0.0005 per URL fetch.
- What are the downsides of Perplexity MCP Server?
- Answer-tool cost per call not published in advance; new accounts limited to 1 Agent API request per second; no free anonymous tier; needs a key with credit; presets change underneath you with no pin in the MCP tools.
- Who is Perplexity MCP Server best for?
- Developers and analysts who want an assistant to return cited answers and research briefs from the live web.
- How much does Perplexity MCP Server cost for a team or in production?
- 10,000 raw searches a month cost $50, or $10 on Fast Search. Answer tools add model tokens plus $0.0025 per standard web search, $0.001 per fast search and $0.0005 per URL fetch, so a research call that makes a dozen searches costs about three cents before tokens.
- Is Perplexity MCP Server safe for company data?
- API content is not used for training; zero data retention only by written agreement; SOC 2 Type II. The hosted server sees every query, and the tools fetch untrusted web pages, so prompt injection through results is possible; an open issue proposes flagging untrusted sources.
- What are the best Perplexity MCP Server alternatives?
- The closest alternatives we have reviewed are Exa MCP Server, Brave Search MCP Server, Tavily MCP Server and Kagi MCP Server.
Answers checked Oct 22, 2026.
Make Perplexity MCP Server? Show this rating on your site
