JuiceBox - AI Powered Animation

Your AI assistant will happily write you an animation. It will be confident about it. And it will usually feel wrong. AI animation fails for a structural reason, not a knowledge one. Feel is not correctness. A tween that compiles, runs, and moves the object exactly where you asked can still read as dead, and the distance between dead and satisfying is an asymmetric curve and about four numbers. Your agent cannot judge those numbers, because feel is judged by watching and your agent cannot watch. It is not missing information. It is missing a sense. So it guesses, and then you pay for the guess. You say "it feels off," because that is the only vocabulary any of us has for this. Now it re-guesses a value it had no principled way to pick the first time, and you play the scene again. Every correction is a full round trip. Code-first tween libraries do not help: the numbers still live in code, the judgment still happens in a channel your agent does not have, and your only recourse is still to describe the gap back to it in words. JuiceBox LLM improves the guess first. A lerp is a blank primitive: every quality of the motion has to be built out of raw numbers, which are very difficult to convert into feel by themselves. JuiceBox hands it concepts instead. A Swing already knows what a pendulum does. A Follow already knows how something chases a moving target. An easing is a named shape rather than four floats to invent. Choosing between five effect families is a judgment an agent is genuinely good at, and the choice carries the feel along with it, so the first attempt lands a lot closer than anything assembled out of lerps. It still cannot see the result. It just no longer has to invent the feel from nothing. Then it moves the judgment for whatever is left. Your agent builds the structure — the right effects, the right sequence, the right targets, the right easings — as a real JuiceBox animation asset. You open it in the graph editor and drag the overshoot until it looks right. You can modify the values live as the animation is running. The agent does the part it is good at, and the person who can actually see the screen does the part that needs eyes. This is AI animation you can edit, and it is not character animation. UI, cameras, GameObjects, effects, and gameplay feedback: the panel that slides in, the hit that flashes, the pickup that floats and then flies to the player. Generated clips and mocap libraries do not touch that layer. It is the one AI tooling in Unity has skipped. Setup is simple: - Import the package. A setup window opens on first launch. - Click "Set up agent access" to set up access and/or connect Unity's CLI or MCP server. - Describe the animation you want. Not a chatbot bolted into the editor. The package publishes JuiceBox's entire animation vocabulary as typed, schema validated agent tools, so an AI powered workflow acts on your project directly instead of guessing at an API it has never seen. It is generative AI aimed at something narrow enough to be reliable: artificial intelligence that authors animation data, in the Unity editor, with no AI code in your build. No API key. No subscription. No credits. No generation limits. JuiceBox LLM never calls a model itself, so there is nothing for it to bill you for and no key for it to ask you for. This package contains is no account to create, no service to sign in to, and no monthly allowance to run out of halfway through a scene. It works with the AI tools you already use. JuiceBox LLM is a language that your model uses to set up animations, so it is model agnostic and tied to no single AI plugin: Claude Code, Cursor, GitHub Copilot, Claude Desktop, Unity's own AI Assistant, ChatGPT and other GPT class models, Gemini, or a local LLM served through Ollama. Natural language in, working animation out. Describe what you want in the AI chat you already have open and the integration builds a real sequence, then keeps editing it conversationally as you refine. This i