oxframe

Explainers, code walkthroughs, charts and release videos from code. Rendered natively — no Chrome, no Node — deterministic, and checked in under a second with a report that points at the line to fix. MIT or Apache-2.0: no seats, no per-render fees.

npx skills add oxygenframe/oxframe-sdk

Installs the oxframe skills into Claude Code, Codex, Cursor, Copilot and others. Then ask it for a video. Agents and skills

Every frame drawn by oxframe. View the code of this video

How it works with agents

Prompt, then check, then render. The agent reads the picture and the report, not only the compiler.

Prompt

Check

{
  "duration": 2.7,
  "issues": [{
    "rule": "text_overlap",
    "severity": "error",
    "time": [1.4, 2.3],
    "node": "line",
    "source": "src/title.rs:218",
    "hint": "Text overlaps title. Move one of them."
  }]
}

Render

Numbers

oxframe 0.23.9, measured 2026-10-26 on Intel(R) Core(TM) Ultra 7 255H, 16 threads, rustc 1.97.1, Linux. One scene: a title card, a bullet list and an end card, about 28 seconds. oxframe was already built in the same target directory.

0.58s

One-line rebuild

cargo​ build after changing a single string, with oxframe already compiled.

0.28s

Warm check

oxframe​ check when that build is already current.

0.82s

Check after an edit

oxframe​ check right after a change, the build included.

5.09s

Render of a 27.52s video

oxframe​ render on the CPU with 8 parallel jobs (12.23s with one). The two files are byte for byte the same.

98%

Agent success

Claude Code passed 46 of 47 reference prompts: the scene built, oxframe​ check was clean and it had what the prompt asked for (oxframe 0.23.4, 2026-10-25).

This is self-reported: a prompt passes when the scene builds, oxframe's own oxframe​ check reports no error and the scene meets the prompt's expectations (the nodes it should use, the aspect ratio, a minimum length), run with one agent (Claude Code), 20 turns per prompt on average. A comparison across tools, judged independently, is still to come — the prompt set and the runner are open in the CLI.

You don't need to know Rust. Your agent does.

Themes and components carry the look. This scene is the whole program: a title, three steps and an end card.

Open the component example

use oxframe::prelude::*;

#[oxframe::scene]
async fn intro(view: View) {
    let title = view.add(TitleCard::new("Binary search").subtitle("O(log n)"));
    title.enter().await;
    let steps = view.add(BulletList::new(["Take the middle", "Compare", "Drop half"]));
    steps.enter().await;
    for _ in 0..steps.len() {
        steps.next().await;
    }
    let end = view.add(EndCard::new("That's the search"));
    end.enter().await;
}

fn main() {
    Project::new("intro").theme(Theme::dark()).scene(intro()).main();
}

Features

oxframe check

A storyboard and a JSON report after every change. Errors name the rule, the time, the node and the line.

MCP server

oxframe mcp gives an agent check, inspect, a single frame and the docs for this version.

Themes

Color, type and motion come from tokens, so a scene looks finished before anyone picks a hex value.

Components

Title cards, lists, code, diagrams, callouts and end cards. Add them like any other node.

16:9, 9:16 and 1:1

One scene, three frames. Relative placement follows the safe area of each shape.

Deterministic

Fixed time steps and a seeded random source: the same code renders the same frames every time — byte-for-byte on the CPU renderer.

Edit timing without recompiling

Markers, eases and tunables live beside the code. Drag them in the editor.

Fast preview

Save, and the editor reconnects to the new build. Scrubbing re-simulates without drawing every frame.

CPU and GPU

tiny-skia on the CPU by default, byte for byte the same on any machine; vello on wgpu with --gpu.

Video out

MP4, WebM with alpha, ProRes 4444, GIF or a PNG sequence, with the soundtrack muxed in.

Inspired by Motion Canvas: scenes are coroutines, and an await is a point on the timeline. oxframe runs that model in Rust, with a native renderer and a check an agent can read.

How it compares

	oxframe	HyperFrames	fframes	Remotion
Language	Rust	HTML, CSS, JS	Rust	React
Where it renders	Native CPU and GPU	Headless Chrome, from HTML	GPU (Skia) or CPU, from SVG	Headless browser
Live editor	Timeline, markers, tunables	Studio and a web playground	GPU preview window and a WebAssembly editor	Preview studio
An agent can see the picture	oxframe check: storyboard and a report with line anchors, plus MCP	check --snapshots, snapshot --review/--against, compare	A skill, plus frame snapshots	Agent Skills; render stills and read them

The full comparison, with Motion Canvas and Manim too, and guides for moving from Remotion, Motion Canvas and Manim.

Works with your agent

Claude Code

oxframe​ init writes a skill and AGENTS.md. The docs are on the agents page.

oxframe init my-video

Cursor

The same AGENTS.md is what Cursor reads. Refresh it after an upgrade with oxframe​ agents.

oxframe agents

Codex

Codex reads AGENTS.md from the project before it edits. oxframe​ init writes that file.

oxframe init my-video

ChatGPT and other chat apps

No plugin (Codex has one, above). Connect mcp.oxframe.dev, or give the model the project's AGENTS.md, or the docs of this version at /llms.txt.

MCP

Any client that speaks MCP can run oxframe​ mcp from the project directory; the docs and a scene check are also hosted at mcp.oxframe.dev. Setting up the MCP server.

oxframe mcp

llms.txt

The docs of this version as plain text, for tools that read the web.

/llms.txt and /llms-full.txt