Correction, 2026-10-01: this page said a custom connector in claude.ai needs Claude Pro, Team or Enterprise. Anthropic's help center lists custom connectors on every plan, Free included (Free allows one), under a different menu path. It also counted the 8frame connector's tools as models, showed an assistant upscaling a clip after animating it (8frame has no video upscaler; upscaling is for images), and said the server routes a described shot to the right engine and returns the clip. In fact each tool runs one model family, chosen per call, and results come back as file links that do not land on the canvas. It said OpenAI and Google added MCP support on dates and in words we could not confirm on a first-party page, and called MCP the industry's default way to give AI systems tools, which we have not measured. Those lines now follow the MCP documentation, and the page has a FAQ and a list of sources.
MCP (Model Context Protocol) is an open standard that lets AI assistants connect to external tools and data sources through one common interface, so a model can take actions instead of only generating text. Anthropic introduced and open-sourced it on November 25, 2024, as "a new standard for connecting AI assistants to the systems where data lives." The MCP documentation now lists support in AI assistants such as Claude and ChatGPT and in development tools such as Visual Studio Code and Cursor.
The problem MCP solves is combinatorial. Before it, every AI application had to build a custom integration for every tool it wanted to use: one for your files, one for your database, one for an image generator, one for each API. In Anthropic's words, "every new data source requires its own custom implementation." That's an M-times-N problem. MCP turns it into M-plus-N: a tool exposes itself once as an MCP server, any MCP-capable AI app acts as a client, and they speak the same protocol. Build the integration once, and any application that supports the standard can use it.
That standardization is why the protocol spread beyond Anthropic's own apps. OpenAI's developer documentation describes connecting remote MCP servers to ChatGPT and to its API, and the MCP site sums up the promise as "build once and integrate everywhere." Before you build for a particular host, check its documentation for the parts of MCP it supports, such as sign-in for remote servers.
How MCP works
MCP has three roles. The host is the AI application the user interacts with (a chat app, a coding editor, an agent). The host creates one client for each server it connects to, and each client keeps a dedicated connection to its server. The server is the program that exposes a specific capability: your calendar, a code repository, a video generator. One host can connect to many servers at once. A server can run locally on your machine or remotely over the web; for remote servers that need you to sign in, the MCP documentation recommends OAuth.
Servers expose their capabilities in three standard shapes, which the documentation calls primitives. Tools are actions the model can call (generate an image, run a query, send a message). Resources are data the model can read (files, records). Prompts are reusable templates the server offers. The model doesn't need to know how any of it is implemented; it sees a tool with a name, a description and a schema for its inputs, calls it, and the server does the work and returns a result.
The flow in practice: you ask the assistant to do something, the model decides a tool call is needed, the host sends that call to the right server through its client, the server executes and returns a result, and the model uses the result to continue. Because the protocol is standard, a server you connect to Claude can also be connected to other MCP-capable hosts without rewriting it, as long as each host supports what the server needs. That's the whole value: write the connector once.
When you use MCP
You use MCP whenever you want an AI system to act rather than just answer. A plain model call generates text. An MCP-connected model can read your data and take actions through real tools. Common situations:
- Connecting an assistant to your systems. Give a chat assistant access to your files, database, or internal tools without building a bespoke integration for each.
- Building agents. An AI agent needs tools to do anything beyond talk. MCP is a standard way to hand it those tools, so you're not maintaining a separate client per capability.
- Adding a capability to many hosts at once. If you build a tool and want it usable from Claude, an IDE, and a custom agent, one MCP server can serve all three, provided each host supports the features the server uses.
The signal that you want MCP is repetition: if you'd otherwise write the same kind of integration more than once, or want one tool to work across multiple AI apps, the standard is doing exactly the job it was designed for.
Examples
A creative generation server (8frame). 8frame runs a live MCP server at https://mcp.8frame.co/mcp. As of 2026-10-01 it offers 8 image-generation tools, 9 video tools and 2 music tools, plus tools for image editing, background removal and image upscaling. Each tool runs one model family (the Veo tool, for example, covers Veo 3.1 Standard, Fast and Lite), and two "pick a model" tools list the options with their credit costs. Once you connect it, an assistant like Claude can generate a product still, upscale it and animate it into a clip, all through tool calls, without you opening the app.
To connect it in claude.ai, add it as a custom connector under Customize → Connectors → + Add → Add custom connector. Custom connectors work on every Claude plan, Free included (Free allows one); on Team and Enterprise an Owner adds them for the organization. In Claude Code, run claude mcp add --transport http 8frame https://mcp.8frame.co/mcp, then /mcp to sign in. In Cursor, add the URL under mcpServers in ~/.cursor/mcp.json. Either way, you sign in to your 8frame account in the browser, and generations are charged to that account's credits.
The assistant picks a tool for each call, sends the prompt, and gets the result back as a file link; results don't land on your 8frame canvas. The connector also has fewer settings than the canvas. Nano Banana Pro and Kling v3 Pro are canvas-only: the connector's Nano Banana tool is the 6-credit Classic model, and its Kling tool runs Kling v3 Standard, which needs a start image. A Veo call that names no tier runs Veo 3.1 Standard at 448 credits for 8 seconds with sound, while the Veo tool on the canvas opens on Fast (168), so name the tier you want.
A code repository server. A developer tool can expose a GitHub-style server with tools like "read file," "open pull request," and "search issues." Anthropic's launch shipped pre-built servers for GitHub, Git, Google Drive, Slack, Postgres and Puppeteer. Connected to a coding assistant, the model can read the codebase and take actions on the repo through those tools, and the same server, built once, can be connected to other MCP hosts the developer uses. This is the everyday pattern: a system that already has an API wraps it as an MCP server, and MCP-capable assistants can reach it without a custom integration each.
Both examples share the same structure: a server exposing named tools, a host with a client that connects to it, and a model that calls the tools to get real work done. Whether the capability is generating video or editing code, the protocol underneath is identical.
Related concepts
An AI agent is the most common consumer of MCP. Agents need tools to act on the world, and MCP is a standard interface that supplies them, which is why the two concepts are usually discussed together. For the 8frame side of that, see how to add video generation to an AI agent and how to generate AI video from Claude.
An AI workflow is the other side of the coin: where MCP standardizes how a model reaches a single tool, a workflow chains multiple generation steps into a repeatable pipeline. On a canvas you build the workflow; through MCP you can run many of the same models from a chat or an agent, with fewer settings, and get the results back as files.
FAQ
Who created MCP? Anthropic introduced and open-sourced it on November 25, 2024, together with a specification, SDKs and a set of pre-built servers.
Is MCP only for Claude? No. It is an open protocol, and the MCP documentation lists support in ChatGPT, Visual Studio Code and Cursor as well as Claude.
What is the difference between an MCP host, client and server? The host is the AI app you use, the server is the program that offers tools or data, and the client is the part of the host that keeps one connection to one server.
Do I need a paid Claude plan to add a custom MCP connector? No. Anthropic's help center lists custom connectors on Free, Pro, Max, Team and Enterprise; Free is limited to one.
Do 8frame results appear on my canvas when I generate through MCP? No. The connector returns each result as a file link. Generations are charged to the credits of the 8frame account you signed in with.
Sources
- MCP introduced and open-sourced on November 25, 2024, the launch wording and the pre-built servers: Introducing the Model Context Protocol, checked 2026-10-01
- Definition and support in Claude, ChatGPT, Visual Studio Code and Cursor ("build once and integrate everywhere"): What is the Model Context Protocol (MCP)?, checked 2026-10-01
- Host, client and server roles, tools, resources and prompts, local and remote servers, OAuth: MCP architecture overview, checked 2026-10-01
- Remote MCP servers in ChatGPT and the OpenAI API: OpenAI developer docs, MCP, checked 2026-10-01
- Claude custom connectors (plans, the Free limit, who adds them, menu path): Get started with custom connectors using remote MCP, checked 2026-10-01
- Claude Code command and
/mcpsign-in: Claude Code docs, MCP, checked 2026-10-01 - Cursor config file and OAuth: Cursor docs, MCP, checked 2026-10-01
- 8frame MCP setup and sign-in: 8frame.co/mcp, checked 2026-10-01
- 8frame connector tools, models, defaults, prices and file-link results: 8frame MCP server code, read 2026-10-01
- Canvas Veo default (Veo 3.1 Fast, 168 credits): the 8frame tool registry, 2026-09-30
Want to see MCP in action? Connect the 8frame MCP server to Claude and generate video from a conversation, or explore the canvas and workflow library for the shots you direct by hand.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.