MCP server

oxframe has two MCP servers. oxframe mcp runs on your machine and lets an agent check, inspect and render the video it is writing. https://mcp.oxframe.dev/mcp is hosted, needs no install and serves the documentation to agents that cannot run commands.

	Local: `oxframe mcp`	Hosted: mcp.oxframe.dev
Transport	stdio	Streamable HTTP, `https://mcp.oxframe.dev/mcp`
Install	The oxframe CLI, Rust, ffmpeg	Nothing
Sign-in	None	None
Tools	Check, storyboard, frames, inspector, render, docs	Docs search, pages, components, examples
Good for	Claude Code, Codex, Cursor, VS Code: agents that edit files	ChatGPT, Claude.ai, web IDEs: agents that only read
MCP Registry	`io.github.oxygenframe/oxframe`	—

The local server, from scratch

1. Install the oxframe command (Linux and macOS; Windows and other ways are in Install and first project):

curl -fsSL https://github.com/oxygenframe/oxframe-cli/releases/latest/download/install.sh | sh
# or: brew install oxygenframe/tap/oxframe
# or: cargo install oxframe-cli
oxframe doctor   # Rust, ffmpeg and a C linker in place?

2. Create a project. The server works on the project in the directory it starts in.

oxframe init my-video && cd my-video

3. Add the server to your client, from inside that directory.

Claude Code

claude mcp add oxframe -- oxframe mcp
# or, shared with everyone who clones the project (writes .mcp.json):
claude mcp add --scope project oxframe -- oxframe mcp

The oxframe plugin adds the same server together with the skills: /plugin marketplace add oxygenframe/oxframe-sdk, then /plugin install oxframe@oxframe.

Codex

codex mcp add oxframe -- oxframe mcp

Or in ~/.codex/config.toml:

[mcp_servers.oxframe]
command = "oxframe"
args = ["mcp"]

Cursor

.cursor/mcp.json in the project, then open the project folder in Cursor:

{
  "mcpServers": {
    "oxframe": { "command": "oxframe", "args": ["mcp"] }
  }
}

VS Code (GitHub Copilot)

.vscode/mcp.json in the project:

{
  "servers": {
    "oxframe": { "type": "stdio", "command": "oxframe", "args": ["mcp"] }
  }
}

Any other client

The command is oxframe with the argument mcp, started in the project directory. Messages are newline-delimited JSON-RPC on stdin and stdout; a Content-Length frame is accepted too. Logs go to stderr.

Tools of the local server

`check`	Compiles the project and runs the visual check. Returns `report.json` (each issue with a rule, a time range, the node and `file:line`), the storyboard path and hints for common compile errors. `frames` also writes every sampled frame; `against` compares with an earlier check and writes diff images.
`storyboard`	The same check, answered with the storyboard path and a short summary.
`render_frame`	One frame as a PNG: `at` in seconds, optional `scene` and `size` (default 960×540).
`inspect`	The scene tree at one time: kind, name, bounds, text and `file:line` of each node.
`render_graph`	The render passes at one time, in order, with the shader of each effect pass.
`list_components`	The explainer components (TitleCard, BulletList, CodeExplainer, FlowDiagram and the rest) with a one-line example each.
`docs_search`	Searches the docs of the CLI's own version.
`render`	Builds in release mode and renders the video: `scene`, `aspect` (16:9, 9:16, 1:1 or all), `format`, `size`, `cpu`/`gpu`.

Everything runs on your machine; the local server sends nothing anywhere. A typical turn: the agent edits src/main.rs, calls check, opens the storyboard and the close-ups of each issue, fixes them and checks again until the report is clean, then calls render. Check rules lists what the check looks for.

The hosted docs server

https://mcp.oxframe.dev/mcp speaks Streamable HTTP, needs no sign-in and only reads. It follows the current docs on every site deploy.

`docs_search`	Searches the current docs and returns the best sections with their page URLs.
`get_page`	The whole text of one docs page as Markdown, by URL or path.
`list_components`	The explainer components and a one-line example of each.
`get_example`	The Rust source of an example from the gallery, by name.

# Claude Code
claude mcp add --transport http oxframe-docs https://mcp.oxframe.dev/mcp
# Codex
codex mcp add oxframe-docs --url https://mcp.oxframe.dev/mcp
# Cursor, .cursor/mcp.json
{ "mcpServers": { "oxframe-docs": { "url": "https://mcp.oxframe.dev/mcp" } } }
# VS Code, .vscode/mcp.json
{ "servers": { "oxframe-docs": { "type": "http", "url": "https://mcp.oxframe.dev/mcp" } } }

In ChatGPT and Claude.ai, add it as a custom connector with the URL https://mcp.oxframe.dev/mcp and no authentication. It keeps nothing about you or your project; requests are logged by Cloudflare like any other (Privacy).

Try it by hand

Both servers answer a plain request, which is handy to see that a setup works:

# local, in a project directory
printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' | oxframe mcp

# hosted
curl -s https://mcp.oxframe.dev/mcp -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

When a tool fails

"no Cargo.toml here or above": the client started the server outside the project. Open the project folder in the editor, or start the client from it.
The client cannot find oxframe: GUI apps may not see your shell's PATH. Use the full path from which oxframe as the command.
A compile error: check returns the compiler text and hints; the agent fixes the code and calls it again. More in Troubleshooting.

How the instructions in each project and the skills fit with the server: Working with AI agents.

Edit this page on GitHub