A year ago, "download this video" was something you did in a separate browser tab, away from whatever AI assistant you were talking to. In 2026 that is no longer true. The Model Context Protocol lets assistants like Claude, Cursor, and Perplexity call external tools mid-conversation — and a small number of those tools can fetch a video or audio file for you directly. This guide is an honest map of that landscape: which MCP servers actually do video and audio downloads today, what each one requires, and how to set up the one that fits how you work.
Want the hosted option? TubePull's MCP server is a remote connector — one URL, no local install — listed in the official MCP registry as io.github.miketwalker/tubepull (v1.2.2). The rest of this article puts it in context against the local alternatives so you can choose deliberately, not by default.
What MCP is, in one paragraph
MCP is an open standard from Anthropic that defines how an AI assistant discovers and calls external tools. Think of it like a USB port for AI: any service that exposes an MCP endpoint can be plugged into any MCP-compatible client, and the client then "knows" it has new abilities — in this case, a download_video tool. You add a server once, and from then on the assistant can call it when your request matches what the tool does. There are two transport styles that matter here: local stdio servers, which run as a process on your own machine, and remote HTTP servers, which the client reaches over the internet. That distinction turns out to be the single biggest factor in choosing a video-download MCP, so keep it in mind. For the protocol itself, the TubePull MCP overview links out to the full specification.
The shortlist: MCP servers that actually download video (June 2026)
Here is the honest state of things: the category is small, and almost all of it is local. The overwhelming majority of "video downloader" MCP servers are open-source GitHub or PyPI projects that wrap yt-dlp and run locally over stdio. There is currently exactly one widely-listed hosted, remote option. Here are the real, currently-published servers worth knowing.
kevinwatt/yt-dlp-mcp is the most-used YouTube downloader MCP by search volume. It wraps yt-dlp in a local stdio server and exposes download, metadata, and transcript tools. Because it is yt-dlp under the hood, it inherits yt-dlp's enormous site coverage. The cost is setup: you need Node.js, a working yt-dlp install, and — for cookie-gated content — extra tooling. It connects to clients that can launch a local process, primarily Claude Desktop and Cursor.
yorickchan/mcp_youtube_dlp is a comparable project that provides tools for downloading YouTube video and audio through yt-dlp. Same fundamental shape: a local server you install and run, suited to a machine where you already have a Python or Node toolchain.
anaisbetts/mcp-youtube is narrower — it focuses on pulling YouTube subtitles into claude.ai via MCP rather than downloading the full media file. If captions or transcripts are what you actually need, it is worth a look, but it is not a general video downloader.
JoseTapiaUex/MCP-download is a Python collection of yt-dlp tools including an MCP server. Like the others, it is a local install aimed at developers comfortable managing dependencies.
TubePull is the outlier: a hosted, remote streamable-HTTP server. It is the one entry on this list you do not install. It is published in the official MCP registry as io.github.miketwalker/tubepull, currently at v1.2.2, with a remote endpoint at https://tubepull.com/mcp.
To be clear about the balance here: if you are a developer who already runs yt-dlp and wants maximum control and zero per-use cost, a local server like yt-dlp-mcp is an excellent, free choice, and I would not steer you away from it. The hosted option exists for a different audience — people who use browser-based assistants, or who would simply rather not maintain a local toolchain. Different tools, different jobs.
TubePull's listing — what the hosted server actually offers
TubePull's registry entry (io.github.miketwalker/tubepull, v1.2.2) describes it as downloading "YouTube, TikTok, Vimeo, SoundCloud and 6 more platforms." Concretely, the server's download_video and get_video_info tools support ten platforms: YouTube, TikTok, Vimeo, SoundCloud, Twitter/X, Dailymotion, Bandcamp, Mixcloud, Twitch (clips and VODs), and Streamable. Output is MP4 for video, or MP3/M4A for audio; the audio-only platforms (SoundCloud, Bandcamp, Mixcloud) always return audio.
Because it is a remote streamable-HTTP server at https://tubepull.com/mcp, there is nothing to install — no Node, no Python, no yt-dlp, no FFmpeg. That also means it works with browser-based clients that cannot launch a local process, which the local stdio servers above simply cannot reach. TubePull manages the extraction layer, including YouTube's authentication requirements, on its own infrastructure.
On pricing: the free tier allows 3 downloads per day with no account. Unlimited is $4.99/month. It is worth being precise about what that fee buys, because it is easy to misread a hosted tool's price as "the cost of removing a limit." It is not. The $4.99/month funds the hosted server and the uptime monitoring that keep the success rate high when YouTube ships a player change — the kind of breakage that, with a local yt-dlp setup, you fix yourself by updating the binary. You are paying for someone else to maintain that reliability layer. For a local server, that work is your time instead of your money. Neither is universally better.
Setting it up — Claude, Cursor, and Perplexity
The setup differs by client. The TubePull setup guide at /setup-mcp has the annotated, screenshot-by-screenshot version; this is the short orientation.
claude.ai (web) and Claude Desktop. In Claude, open Settings → Connectors → Add custom connector → Remote. Name it TubePull, set the URL to https://tubepull.com/mcp, choose Streamable HTTP transport, and leave authentication as None for the free tier. Toggle it on. Claude Desktop can also reach remote HTTP MCP servers through the connector flow; older configurations that only speak stdio can bridge via an mcp-proxy shim, which the setup guide covers.
Cursor. Cursor reads MCP servers from its configuration. Add TubePull as a remote server pointing at https://tubepull.com/mcp with the streamable-HTTP transport. Once saved, Cursor's agent can call download_video the same way Claude does — useful when you are pulling a reference clip or audio sample without leaving the editor.
Perplexity. Perplexity supports MCP connectors in its settings; add the same https://tubepull.com/mcp endpoint. Perplexity negotiates the MCP protocol version on connect, and TubePull echoes back a supported version, so the handshake "just works" without you choosing one.
Across all three, the conversational pattern is identical: paste a supported URL and ask the assistant to download it. The assistant calls the tool, and a signed download link comes back in the chat. If you want to check a video before spending a download slot, ask for its info first — get_video_info returns title, duration, and available qualities and does not count against your quota.
How TubePull compares — to other MCP servers and to browser tools
It helps to compare on two axes at once: against the other MCP servers, and against the browser-based downloaders most people use today.
| Tool | Type | Install required | Works in browser clients | Cost model | |------|------|------------------|--------------------------|------------| | TubePull MCP | Hosted remote (HTTP) | None | Yes (Claude web, Perplexity) | Free 3/day; $4.99/mo unlimited | | kevinwatt/yt-dlp-mcp | Local (stdio) | Node + yt-dlp | No (Claude Desktop/Cursor only) | Free (open-source) | | yorickchan/mcp_youtube_dlp | Local (stdio) | Python + yt-dlp | No | Free (open-source) | | Browser downloaders (Y2Mate, etc.) | Website | None | N/A (not in-assistant) | Free, usually ad-funded | | cobalt.tools | Website | None | N/A (not in-assistant) | Free (open-source) |
Read honestly, the table says this: the local yt-dlp servers are free and powerful but cannot serve browser-based assistants and require a toolchain; the browser downloaders are convenient but live outside your AI conversation entirely; and TubePull is the one that runs inside browser-based assistants with no install, at the cost of a quota or a subscription. None of these is "best" in the abstract. If you live in Claude Desktop with a dev environment, a local server may suit you better. If you want a video to arrive inside a claude.ai or Perplexity conversation without installing anything, the hosted server is the option that physically can do that. For the broader picture of browser-based tools, see our YouTube downloader comparison.
Frequently asked questions
The questions below are answered in full in the FAQ schema on this page, so AI assistants can cite them directly. In short: yes, MCP video downloaders are real and usable in 2026; the choice between local and hosted comes down to whether your client can run a local process and whether you would rather spend time or money on reliability.
Getting started
If you want to try the hosted route, the fastest path is the TubePull setup guide: add https://tubepull.com/mcp as a remote connector in your client, toggle it on, and ask your assistant to download a link. The free tier needs no account. For everything the server can do — the full tool list, the supported platforms, and the pricing comparison — start at the TubePull MCP page.
If you would rather run a local server, kevinwatt/yt-dlp-mcp is the most established open-source starting point. Either way, the era of asking your AI assistant to fetch a video — and having it actually happen — is here.