| name | web-research |
| description | Use the Perplexity MCP for any task that needs fresh / external / web-authoritative information. Triggers on requests like "research X", "look up Y", "what's happening with Z", "fact-check this", "find recent news about...", "competitive analysis of...", "what are customers saying about...", "is this still true?", "ask Perplexity...", "deep research on...", "do a teardown of [company]", or any question that benefits from cited live-web answers rather than knowledge from training. Choose the right Perplexity tool — quick lookup → perplexity_search; synthesized answer → perplexity_ask; multi-step analysis → perplexity_reason (Pro); long-form report → perplexity_research (Pro); ASI / heavy computation → perplexity_compute (Pro/Max). When Pro tools are unavailable on a Free account, chain multiple perplexity_search calls and synthesize manually — the source-finding works at every tier. |
Perplexity MCP — when and how to use it
Claude has a Perplexity MCP server installed that hits the live web (often through a logged-in Perplexity session), returns citations, and can run multi-step research. Use it whenever the user is asking about something that benefits from fresh data — current events, real customer reviews, recent product launches, today's pricing, who-acquired-whom news, this-quarter's funding rounds.
Don't use Perplexity for things Claude already knows reliably from training (general concepts, historical events, well-established facts). Don't use it for code, math, internal reasoning, or anything that doesn't need the open web.
Plan tiers (Free / Pro / Max)
The Perplexity MCP routes to different models depending on the user's plan tier. Some tools are completely gated. Plan-tier limits change; verify at https://www.perplexity.ai/pricing.
| Capability | Free | Pro | Max |
|---|
perplexity_search | ✅ | ✅ | ✅ |
perplexity_ask | ✅ (force model: "turbo") | ✅ | ✅ |
perplexity_reason | ❌ | ✅ | ✅ |
perplexity_research | ❌ | ✅ | ✅ |
perplexity_compute (ASI / Computer mode) | ❌ | partial | ✅ |
perplexity_models, perplexity_login, perplexity_doctor | ✅ | ✅ | ✅ |
perplexity_retrieve, perplexity_get_research, perplexity_list_researches | ✅ | ✅ | ✅ |
perplexity_export, perplexity_sync_cloud, perplexity_hydrate_cloud_entry | ✅ | ✅ | ✅ |
If you don't know the user's tier, attempt the Pro tool. If it fails with requires an authenticated Pro account, fall back to the free-tier pattern (see below). Don't ask the user to upgrade.
Tool routing — pick the right tool for the task
| User asks for… | Use this tool | Notes |
|---|
| Quick fact, "what's X?", "find me the URL of…" | perplexity_search | Fastest. Free-tier safe. |
| Synthesized answer, "ask Perplexity…", "summarize what people are saying" | perplexity_ask | Pass model: "turbo" if Free. Returns follow_up_context you can pass back for follow-up questions. |
| Multi-step reasoning, "compare X and Y", "reason through…" | perplexity_reason | Pro. Falls back to a chain of perplexity_search + manual synthesis on Free. |
| Long-form research, "do a teardown of…", "deep research on…" | perplexity_research | Pro. Same fallback. |
| ASI / heavy computation, "model the cost of…", code-style analysis | perplexity_compute | Pro/Max. No good fallback — say so to the user. |
| List models available on the user's account | perplexity_models | Useful for diagnostics. |
| Re-fetch a Perplexity thread by URL or ID | perplexity_retrieve | When the user pastes perplexity.ai/search/<id>. |
| Fetch a specific saved research result | perplexity_get_research | Pair with perplexity_list_researches to discover IDs. |
| List previously saved research threads | perplexity_list_researches | Useful when the user references past work. |
| Export a research result to markdown | perplexity_export | Good for delivering to the user as a file. |
| Sync local state with cloud (cross-device consistency) | perplexity_sync_cloud | Rarely needed in normal chat flows. |
| Re-pull a single cloud entry locally | perplexity_hydrate_cloud_entry | Edge case — only if the user explicitly mentions a cloud-only entry. |
| Diagnose Perplexity install issues | perplexity_doctor | Pass probe: true to also do a live search probe. |
| Trigger interactive login | perplexity_login | Returns instructions only — login is interactive (email + OTP) and must be done from CLI or extension dashboard. |
Sources picker — pick deliberately
Search-style tools accept sources: ["web", "scholar", "social"].
web (default) — general web search; use for nearly everything.
scholar — academic / peer-reviewed; use when the user wants research evidence.
social — Reddit, forums, social media; use for "what are customers actually saying" / sentiment.
Combinations are fine: ["web", "social"] is a great default for product or company research where you want both vendor pages and community discussion.
Free-tier fallback (Pro-tool error)
If perplexity_reason or perplexity_research returns requires an authenticated Pro account:
- Don't give up.
- Decompose the question into 4–6 focused sub-questions.
- Run those as parallel
perplexity_search calls (which work at every tier).
- Synthesize the answer yourself using the URLs each search returned. Lead with the most material finding (acquisitions, leadership exits, big funding rounds, etc.) — those reframe everything else.
- Cite specific URLs in the answer.
The free-tier turbo model produces shorter / more outline-style synthesis, but source-finding is identical at every tier. Free users can get equivalent depth — they just have to chain calls.
Patterns
Pattern: company / competitor teardown
Don't try to do this in one call. Run focused parallel searches:
search 1: "[Company] background — founders, founding year, HQ, funding rounds, valuation"
search 2: "[Company] product — what they sell, pricing model, native vs integrated, architecture"
search 3: "[Company] customer reviews — G2, Capterra, TrustRadius, Reddit praise and complaints"
search 4: "[Company] customer logos — named customers, target segment, industry concentration"
search 5: "[Company] vs [main competitor] — where they win, where they lose"
search 6: "[Company] recent news 2024-2026 — launches, acquisitions, leadership changes, layoffs"
Then synthesize. Lead with the most surprising / material finding. End with "what this means" tying findings to whatever the user is actually deciding.
Pattern: fact-check
Quote the specific claim as the search query, with sources: ["web"]. Inspect citations — if they confirm, cite them; if they contradict, lead with that. Don't hedge if the evidence is clear.
Pattern: customer sentiment / "what are people saying"
perplexity_search with sources: ["web", "social"]. Always include social for sentiment work — Reddit and forum discussion is where the actual user voice lives.
Pattern: current-events question
perplexity_search with sources: ["web"]. The MCP automatically uses fresh web indexes. Add a time window in the query ("recent", "this week", "2026") if relevant.
Pattern: user pasted a URL
If the user pastes a URL and asks about its content, prefer fetching directly when you have a fetch tool (e.g. WebFetch). Perplexity is for discovering URLs you didn't have, not for re-reading ones you did.
Pattern: follow-up questions on the same thread
Save the follow_up_context blob from a perplexity_ask response. Pass it back to perplexity_ask on the next call to continue the same conversation thread. This is more coherent than starting a new search every time.
Writing good Perplexity queries
- Be specific. "Subskribe customer feedback" returns generic CPQ chatter. "Subskribe customer reviews — G2, Capterra, Reddit, Trailblazer Community" returns actual review-platform results.
- Include the time window when freshness matters: "2024–2026", "recent news", "this quarter".
- Name the exact sources you want in the query body. Perplexity uses them as ranking hints.
- Ask for the output shape you want: "list named customers", "quote real reviews", "compare side-by-side".
Handling the response
Perplexity responses include:
- The synthesized answer text
- A numbered Sources list with URLs
- A Suggested follow-ups list (sometimes useful for deciding what to search next)
- A Full thread URL on perplexity.ai (good to share with the user as the canonical record)
- A
follow_up_context JSON blob — pass back to perplexity_ask for follow-ups in the same thread
When relaying results to the user, always include the cited URLs — citations are the whole point of using Perplexity vs. answering from training. Lead with the answer, then cite.
Don't
- Don't burn searches on things Claude already knows reliably.
- Don't dump raw Perplexity output unfiltered. Synthesize. Cite specific URLs. Make a recommendation.
- Don't make up facts. If a search returns nothing useful, say so — better to admit uncertainty than to fabricate.
- Don't ask the user to upgrade to Pro. The free-tier fallback gets equivalent depth with more searches.