ReelMimic

Why it's ranked here

The interesting part is the process control, not the generation. A stage advances only when the agent has actually written the files that step requires, reviewers never grade their own work, and a shared agent budget puts reviews and fixes ahead of new builds. It is MIT licensed, ships its own environment checker, and the README is candid about limits: one to three and a half hours for a 30 to 60 second video, tested mostly on Windows, and five of the seven engines still young. A careful pipeline design with a heavy footprint.

What's good

New looks are plain Markdown files listing recognition cues, defaults and known pitfalls, so a lesson from one production carries into the next without code. Agent instructions reload before each turn, so a prompt fix reaches running jobs with no restart. At approval the drawing engine is copied into the project, so improving a skill never changes a job mid-flight. The server listens on the local machine only, prompts reach the agent through standard input rather than the command line, and file serving is confined to the project folder. Fetched assets are logged with source, author and license, and lyrics come only from text you supply.

Tradeoffs

The agents run with wide permissions. Codex defaults to full system access with approvals off, because its sandbox cannot launch Chrome, and Claude Code gets shell, file edits and web access. Every minute of work counts against your AI plan's usage. The stack is heavy: Node 22.18 or newer, Python 3.10, FFmpeg, Chrome, plus OpenCV and librosa. The agent instructions are written in Traditional Chinese, and they tell agents to look up and reproduce characters the user names, such as mascots or cartoon figures, which is looser than the README's original-characters rule. Clearing rights to the result is on you.

How to use it well

Best for someone who wants short stylized animation, such as a music video, an explainer or a narrated comic, and already pays for Claude Code or Codex. Run it on a machine you can spare for hours with at least five gigabytes free. Spend your attention on the plan and style frames before approving: that is the cheap point to change direction. If full system access worries you, the restricted Codex sandbox still handles planning but cannot render. It does not make live-action footage or real people, and 3D is paused. Posting to social platforms is a separate opt-in command that defaults to a dry run.

Observed

License

MIT for the app code; bundled third-party skills keep their own Apache-2.0 or MIT licenses

Language

TypeScript server and React web app; Python analysis scripts

Interface

Local web app with a REST and Server-Sent Events API on port 4318

Agent backends

Claude Code CLI or Codex CLI, run headless on the user's own login

Runtime requirements

Node 22.18+, Python 3.10+, FFmpeg, Chrome

Platform support

Tested mostly on Windows; macOS and Linux stated as less tested

Output

2D animation only, seven drawing engines

Network exposure

Server binds to 127.0.0.1 only

Codex default

Full-access sandbox with approvals disabled, overridable by setting

Setup

Install script plus a built-in environment check command

Tests

Server test files and a CI workflow present in the repository tree