People increasingly want their AI assistant to do the whole job, not just talk about it. "Download this YouTube video for me" is one of the first things they try — and the results vary wildly. Some assistants flat-out refuse on Terms-of-Service grounds. Some try, fetch metadata, and fail to produce a file. And a few actually return a downloadable video right in the conversation. This post tests all six major assistants on the same task and reports exactly what each one can and cannot do as of June 2026.
I build TubePull, whose Model Context Protocol (MCP) server is one of the mechanisms several of these assistants use to complete the task — so this is a disclosed bias. The methodology and per-assistant results below are reproducible; you can run the same test yourself.
Methodology
The test was deliberately simple and identical for every assistant:
- The prompt: "Download this YouTube video for me," followed by a single stable URL.
- The URL: a public-domain clip from the NASA YouTube channel, chosen because NASA content is public domain under 17 U.S.C. § 105 and uploaded by NASA under terms that make downloading uncontroversial. (For the full picture of what is and is not permitted, see our legality guide and the public-domain video archives post.)
- Account tiers: each assistant tested on its paid consumer tier where a paid tier gates the relevant feature (noted per assistant).
- Date of test: June 2026, on the then-current model for each product.
- Definition of success: the assistant returns an actual downloadable file or a working signed download link in the conversation — not a description, not a yt-dlp command for me to run myself, not a metadata summary.
Where an assistant could not do it natively, I recorded whether a supported connector path exists, because in 2026 the connector — specifically MCP — is what actually decides the outcome.
Six assistants, six results
ChatGPT (Plus / Pro / Business) — works via MCP connector
Native: Refuses, citing YouTube's Terms of Service. With browsing: can fetch the title and metadata, but not the file. The path that works: In late 2025 OpenAI shipped full remote MCP connector support in ChatGPT Developer Mode, available on Plus, Pro, Business, Enterprise, and Education accounts. ChatGPT supports remote (HTTPS) MCP servers only — which is exactly what TubePull is. Enable Developer Mode, add https://tubepull.com/mcp as a connector, and "download this YouTube video" returns a signed MP4 link. Verdict: works via MCP (Developer Mode required). Note that Developer Mode grants full read/write tool access, so OpenAI gates it behind a deliberate opt-in.
Claude (claude.ai web + Claude Desktop) — works via MCP connector
Native: Refuses by default. The path that works: Add TubePull's remote MCP connector and the task completes. On claude.ai web, add https://tubepull.com/mcp as a custom connector; ask in the conversation; receive a signed MP4 link. Tested on Claude.ai web (Sonnet 4.6) in June 2026 — four steps end to end. The full walkthrough is in our download YouTube videos with Claude guide. Verdict: works via MCP — and it is the smoothest browser-based setup of the group because claude.ai's connector UI is purpose-built for remote MCP.
Gemini (gemini.google.com + Gemini app) — does not work
Native: Refuses. Connector path: Gemini's consumer apps do not expose a general remote-MCP custom-connector flow the way Claude, ChatGPT, and Perplexity now do, so there is no supported way for an end user to wire in TubePull. Workaround: Gemini will happily write you a Python yt-dlp script to run yourself, but that fails the test (it does not produce a file in the conversation). Verdict: does not work natively. Worth noting: Gemini's real strength is YouTube search and summarization, not downloading — and Google has an obvious business reason never to ship one-click YouTube downloads into its own assistant.
Perplexity (web + Mac/Windows app) — works via MCP connector
Native: Refuses. The path that works: Settings → Connectors → add a custom connector with MCP server URL https://tubepull.com/mcp, then ask in the thread. The Mac app's connector behavior is often more reliable than the web tab. Full steps in our Perplexity MCP guide. Verdict: works via MCP.
Grok (grok.com / X) — does not work
Native: Grok is more permissive in tone than the others and will engage with the request, but it has no consumer-facing remote-MCP custom-connector flow as of June 2026, so it cannot actually fetch and return the file. Workaround: a developer can wire Grok's API up with custom tools, but that is a build project, not something a consumer user can do in the app. Verdict: does not work for end users.
Cursor (cursor.com) — works via MCP connector
This is the pleasant surprise. Cursor is a coding IDE, not a chat app, but it has had first-class MCP support for a while. Add TubePull as an MCP server in Cursor's settings — the config is plain JSON ({"mcpServers": {"tubepull": {"url": "https://tubepull.com/mcp"}}}) — ask in the chat pane, and you get the file. Verdict: works via MCP, and for developers it is arguably the cleanest setup because it is just a JSON entry. The MCP landing page lists Cursor among the supported clients.
What this means
Step back and the pattern is stark. As of June 2026, four assistants can download YouTube via AI — Claude, Perplexity, ChatGPT, and Cursor — and all four use the same mechanism: a remote MCP connector (TubePull's, in this test). None of them does it "natively"; every one of them refuses the bare request and only succeeds once a connector is attached. The protocol is the gate, not the model.
Gemini and Grok are the holdouts, and for understandable reasons. Google has a direct conflict of interest in shipping YouTube downloads into Gemini, and Grok simply has not built a consumer remote-MCP connector flow yet. Both can talk about downloading and even generate scripts, but neither completes the task inside the conversation.
The bigger takeaway is about direction of travel. A year ago, ChatGPT's inclusion in the "works" column would have been impossible — it had no MCP support at all. OpenAI shipping full MCP connectors in late 2025 flipped it. MCP is now an accepted open standard, and the trajectory points one way: more assistants gaining connector support, not fewer. It is reasonable to expect Gemini or Grok to add a consumer connector path before this list is a year old. When they do, the answer to "which AI can download YouTube" will simply be "whichever ones let you attach an MCP server" — which is most of them.
Why none of them do it "natively" — and why that is fine
It is worth being precise about what "refuses natively" means, because it is easy to read it as the assistants being deliberately unhelpful. They are not. A language model on its own has no ability to reach out to the internet, fetch a media stream, transcode it, and hand you a file — that is not a refusal, it is the absence of a capability. What the assistants do have is a cautious default posture toward requests that touch a platform's Terms of Service, so when you ask cold, they explain the caveat instead of pretending to do something they cannot.
The connector model resolves both halves cleanly. The MCP server is the thing with the capability — it does the fetching and transcoding on its own infrastructure — and attaching it is an explicit, user-initiated choice that puts the responsibility for what you download where it belongs: on you. That is why every assistant in the "works" column completes the task the same way. The model supplies the intent and the language understanding; the connector supplies the muscle. Neither is doing anything shady; they are doing exactly what a well-designed tool boundary is supposed to do.
This also explains why "native YouTube download" is unlikely to ever ship in a major assistant. The legal and relationship complexity of baking it directly into the product is high, while exposing a connector framework and letting users attach the tool of their choice is low-risk and already built. The connector path is not a workaround that will be replaced by something better — it is the design these products are converging on.
Comparison table
| Assistant | Native works | MCP connector support | TubePull MCP works | Free tier compatible | Setup time | |-----------|--------------|------------------------|--------------------|----------------------|------------| | ChatGPT | No | Yes (Developer Mode, paid tiers) | Yes | Paid tiers only | ~5 min | | Claude | No | Yes (claude.ai web + Desktop) | Yes | Yes (free 3/day) | ~3 min | | Gemini | No | No consumer flow | No | — | — | | Perplexity | No | Yes (custom connector) | Yes | Yes (free 3/day) | ~4 min | | Grok | No | No consumer flow | No | — | — | | Cursor | No | Yes (JSON config) | Yes | Yes (free 3/day) | ~2 min |
The bottom line
If you want your AI assistant to actually hand you a YouTube file, your real choice is which connector-capable assistant you already use: Claude, Perplexity, ChatGPT, or Cursor — all four work today via a remote MCP server, and the easiest free starting point is Claude or Perplexity with TubePull's free 3-per-day tier. Gemini and Grok cannot do it yet, but the standard that unlocked the other four is spreading, so that gap is likely temporary. For the deeper comparison of hosted MCP options, see our best hosted MCP server for YouTube guide.
Related reading
- How to Download YouTube Videos with Claude — No Install Required
- Download YouTube Videos in Perplexity with MCP
- The Best Hosted MCP Server for YouTube
- Is SSSTik Safe? The $9.99/Week 'Free Trial' Trap Explained (2026)
- Ytmp3.cc and Y2meta Alternatives — 6 Safer YouTube to MP3 Tools (2026)
- TubePull vs NoAdsDL — An Honest Comparison (2026)
- Best Free YouTube Downloader in 2026 — We Tested 10 Tools
- The YouTube Downloader With No Ads (And Why Most Sites Are Full of Them)