MCP is HTTP for the AI world. Here’s the Mux MCP.

Mux is video for developers. A lot of developers are building systems around AI these days, so it stands to reason that Mux is video for AI and we’re determined to push that further. So, to avoid burying the lede: say hello to the first public release of the Mux MCP.

You're probably seeing a lot of talk about MCPs now, but it's for a good reason. There's a lot of hype in the AI world that can make it tough to find the signal in the noise, but MCPs are exciting. In a lot of ways they're the thing we were missing before; they are the signal!

Feelings about MCPs seem to fall into two boats perfectly seen in one single Twitter interaction:

MCPs aren't fundamentally earth shattering magic, but dismissing them altogether misses the interesting bits. I think it's the most exciting to think about the more concrete parallel I see: HTTP. "What is MCP? What can I do with it? Isn't it just [some technical detail]?" are fair questions, but to see why folks are excited about their future, pretend you're asking the same question about HTTP in the early 90s.

HTTP provided a simple request-response protocol that any application could implement. Sometimes those applications were user-facing (browsers) and sometimes they were machine to machine (APIs). MCP does the same for AI-to-service communication, but it's not just a standardized protocol for communication, it's the documentation too. It standardizes how agents discover and call external tools an LLM can use to get a specific task done.

When you build an MCP server, you describe each function (‘tool’) and data source (‘resource’) with a JSON schema. The AI reads those schemas to discover what it can call and how, so you can plug in a database, API, or custom logic through a single interface, without any extra training or glue code.

Alright, let's talk about the Mux MCP

The Mux MCP server exposes Mux's complete API (well...minus some DELETE routes) through a self-documenting interface. Upload videos, create live streams, generate thumbnails, add captions, manage playback policies, dig through engagement data, and monitor video performance. Every Mux API capability becomes available to AI chats and agents without custom integration work.

For example, asking an LLM to analyze this Mux asset:

The implications for agentic workflows like content moderation or content remixing are exciting! You could imagine workflows like the example above where you have agents that know your content policies reviewing new assets as they come in. Or agents that can review pieces of content and find interesting clips to create automatically.

Or maybe you're running a scaled video platform and an ops person might want to use Mux Data to do a deep dive into what in the world caused that spike in errors yesterday. Or...you've got an agent that checks in on data and gives you a summary when something looks off!

For example, analyzing recent incidents:

Why no DELETE?

We haven’t disabled all DELETE routes, but we did disable a few that would be truly destructive and difficult for us to help you roll back if your LLM gets a little too…helpful. The MCP can’t currently delete an Asset or Live Stream, but it could delete a playback ID associated with either.

As the ecosystem and guardrails evolve, we’re open to enabling everything we can. We want to default to as much power as possible, but in our testing for this initial release we found the robots were pretty quick to reach for footguns.

Installing it

We support the recently released Claude Desktop Extensions format, so you can download this DXT file and open it with Claude Desktop to install it. Once it's installed, configure the environment variables you need and you're good to go.

Otherwise, this guide can help you get it installed in a bunch of different tools, but you can also install it in Claude Desktop the "old fashioned" way by updating your MCP config to include something like this:

{
  "mcpServers": {
    "mux": {
      "command": "npx",
      "args": [
        "-y",
        "@mux/mcp@latest",
        "--client=claude",
        "--tools=dynamic"
      ],
      "env": {
        "MUX_TOKEN_ID": "your_token_id",
        "MUX_TOKEN_SECRET": "your_token_secret"
      }
    }
  }
}

This thing is fresh

I mean that both because I was born in the 80s and it's cool, but also that this is new! MCPs get a lot better as you tweak tool descriptions and names to help hint the LLMs towards what they should be using, so we'll be iterating a lot over the next few months.

So, make sure to give us feedback! We especially want to hear if you run into situations where your LLM isn't finding the appropriate tools for your use case, but any paper cuts or issues you run into are extremely valuable.

What's next

• Iterating on the experience
• A hosted MCP server
• A Docs MCP (this MCP is for using Mux, but we've gotten a lot of value out of MCPs that live in your code editor and help you build integrations)
• The stuff you ask for as you use it 😉