Manim Animation for ChatGPT

Turn an explanation into a real animation, then refine it in conversation.

An open-source MCP plugin that renders Manim Community scenes and presents an HTML5 player through MCP Apps. The conversational model authors a bounded scene specification; the service validates, renders, stores, and displays it. No second LLM API call is needed.

Release status: tested developer preview, installed in a personal ChatGPT account. ChatGPT Work selected the tool from an untagged animation request, rendered a real MP4, played it inline, and preserved a follow-up revision. Initial app loads needed the host's Try again control; this is not a permanently resolved issue. The private runtime and temporary HTTPS media tunnel depend on the local machine. Stable production hosting and public directory approval remain outstanding. This project is not affiliated with or endorsed by OpenAI.

Actual ChatGPT Work: the follow-up doubled duration to 11 seconds and changed the final square to orange. The same player also worked in the inline response card.

What it does

• Renders text, safe mathematical notation, sampled graphs, shapes and algorithm diagrams.
• Supports movement, rotation, scaling, transformations, highlighting and paced timelines.
• Produces H.264 MP4 at 480p, 720p or 1080p, up to 90 seconds.
• Preserves scene specifications, identifiers, revision parents and editing instructions.
• Reuses identical renders within an authenticated account.
• Provides play/pause, native scrubbing, speed controls, host-permitted fullscreen, temporary downloads, source specification, loading states and diagnostics.
• Runs every render in a disposable, resource-limited, network-disabled container.

Try: “Animate a hash collision using separate chaining.” Then: “Make it twice as slow.” The model must select the installed plugin and translate the request to the supported scene vocabulary; tool descriptions and the bundled skill guide this choice but cannot guarantee automatic invocation.

Examples

The Fourier example is a finite sine-series amplitude spectrum, not a full numerical transform demonstration. Its conventions and omissions are stated in its caption. The model can construct other scenes using sampled polylines and explicit timelines.

Run locally

Prerequisites: Linux, Python 3.12+, and rootless Podman with cgroups v2. Docker-compatible worker commands are supported, but the verified isolation path is rootless Podman. FFmpeg, fonts and TeX live in the worker image.

git clone https://github.com/OwenPawl/chatgpt-manim.git
cd chatgpt-manim
python3 -m venv .venv
.venv/bin/pip install -r requirements.lock
.venv/bin/pip install -e '.[test]'
podman build -f worker/Dockerfile -t localhost/chatgpt-manim-worker:0.1.0 .
.venv/bin/python scripts/configure_local.py
.venv/bin/python scripts/serve.py

The endpoint is http://127.0.0.1:8787/mcp. Credentials are generated into .env (mode 0600), never printed or committed. Configure your MCP client's bearer credential from MANIM_API_KEY. The portable mcp.json supplies the endpoint; it intentionally contains no secret. A configured MCP client can discover animation_capabilities, render_animation, get_animation and revise_animation.

.venv/bin/pytest -q
MANIM_INTEGRATION=1 .venv/bin/pytest -q tests/test_integration.py
.venv/bin/python scripts/render_examples.py

For a browser smoke test, set MANIM_ENABLE_PREVIEW=1 in .env, restart, and run .venv/bin/python scripts/smoke_remote.py. It writes a 15-minute scoped test-host URL to .agent/preview-url.txt. Open that URL in a browser. The harness uses the same player and documented MCP Apps bridge messages; it is not the real ChatGPT host.

Install in ChatGPT

See installation for the exact supported flow. Current OpenAI documentation instructs testing custom plugins in ChatGPT Work. Work passed explicit and untagged invocation plus revision tests. A separate standard Chat test explicitly invoked the plugin and played the result inline, with host loading issues documented in the verification record. Native automatic animation integration is not claimed.

The current custom-server form supports OAuth, no authentication, or mixed OAuth. It does not accept a fixed bearer key for a public Server URL. Configure OAuth for that route. For the verified private Secure MCP Tunnel route, select Tunnel and No authentication in ChatGPT while tunnel-client supplies the local server bearer credential. The tunnel itself is authenticated and workspace-scoped; the public MCP endpoint remains protected. See the precise configuration below.

Deployment and security

Read deployment, architecture, security, and verification. The service is designed for a single small Linux host: native MCP process plus rootless worker containers, behind a TLS reverse proxy. A server Dockerfile and Compose configuration are included for operators who use a rootless Podman socket. No cloud signup or paid hosting is required for local use.

Model-authored Python is not accepted. Safe declarative input is the intentional first-release boundary. The worker has no credentials, host mounts or network. This is not a general code runner.

Project layout

plugin.json / mcp.json       Portable Agent Plugins package
skills/manim-animation/     Advisory tool-selection and revision guidance
src/chatgpt_manim/           MCP server, auth, persistence, schema and worker launcher
worker/                     Trusted Manim translator and container image
ui/player.html              Standalone MCP Apps player
examples/ / assets/         Scene specs, real MP4s and verified screenshot
scripts/ / tests/           Setup, rendering, smoke checks and automated tests

MIT licensed. Contributions are welcome; see CONTRIBUTING.md.