Perplexity
Official MCPAnswers with citations rather than pages — cheaper in tokens than crawling, but you inherit Perplexity's judgement about what was relevant.
Answers with citations rather than pages — cheaper in tokens than crawling, but you inherit Perplexity's judgement about what was relevant.
The cheapest way to give an agent real web search, and one of very few independent indexes left. Local server, plain API key.
Built for retrieval quality rather than page volume — pair it with Firecrawl when the agent needs to find pages and then read them properly.
The category default: handles JS-rendered pages, returns markdown an LLM can actually use, and the MCP server mirrors the full API.
What you reach for when Playwright MCP can't run locally — same capabilities, hosted, with session recording. You pay per browser-hour.
Widest structured coverage anywhere — Maps, LinkedIn, Amazon and thousands more. Running Actors costs money, so an unsupervised loop is a billing risk.
Community-only, and Apple gives it nothing to build on: no public API, no llms.txt, no official position. Capable projects (99 tools in one), but you carry the maintenance risk.
Classic 'official on paper': Adobe has a connector, but for actual editing you run a community CEP bridge. Treat the official status as marketing, not capability.
Write-only in practice: agents build a draft file, a human opens and exports it. No official server, no editing API, despite CapCut publishing an llms.txt.
The strongest agent story in video editing — official, full timeline read/write, no auth because it is local. Pay the $295 Studio licence or none of it works.
The AI-native answer to the NLE gating problem: free MCP server on 127.0.0.1, no licence wall, full timeline read/write. Editing depth is nowhere near Resolve or Final Cut, and it is macOS 26+ only.
Generation, not editing: agents produce clips and images, they don't cut a timeline. Official, MIT-licensed server and clean API-key auth.
The AI-native shape of this category: official remote MCP, real write actions, no desktop app required — with editing depth well below a true NLE.
Free-plan-friendly with real write access; beta status is the only caveat worth flagging to a team.
Full read/write on every plan — just make sure you are on the v2 endpoint before the beta shutdown in November 2026.
Strong read/write coverage on free plans; the paywall is on AI search and admin controls, not on MCP itself. No bearer tokens — interactive OAuth only.
One of the first official remote servers and still one of the cleanest — full read/write on issues, zero setup friction.
Was among the first five official servers in the wild; read/write is solid, watch the per-plan API complexity budget.
Widest product coverage of any PM server; the gate is rate quota by plan tier, not access.
Biggest install base, worst setup: official but experimental, local-only, read-only, and gated behind Google Cloud auth.
Official server exists, but it sits behind an enterprise contract — fine if you already pay for Pendo, not a starting point.
Solid official read-only server — good for analysis sessions, not for changing tracking.
Official OAuth server with the cleanest auth story in analytics, but analysis-only — agents read charts, they don't change instrumentation.
Read-only, which is right for session data — the value is giving an agent the replay context behind a bug report.