This repository contains the results and source code for a Proof of Concept (POC) benchmark comparing the capabilities of various AI coding assistants and agents when developing interactive and motion-rich frontend applications using a Motion.dev Model Context Protocol (MCP).
The benchmark evaluated four coding agents across distinct Motion.dev-heavy animation prompts (e.g., a MagneticDock, an interactive SVG blob, a scroll-driven cinematic timeline, and a liquid menu). The goal was to observe how each agent utilized the Motion.dev MCP and how that affected their speed and implementation success.
The full benchmark data, execution reports, and a live showcase of the generated applications are hosted on the gh-pages branch of this repository.
The project is organized into top-level directories based on the AI tool used, and sub-directories based on the prompting strategy or development task employed.
- Claude Code (
/claude-code) - Codex (
/codex) - Gemini CLI (
/gemini-cli) - OpenCode Minimax 2.5 (
/opencode-minimax-2.5)
Within each assistant's directory, you will find experiments categorized by the following approaches/tasks:
- Context-Rich: Providing extensive background and related code context.
- Goal-Oriented: Defining the end-state and allowing the agent to determine the path.
- Iterative: Building the solution step-by-step with back-and-forth prompting.
- Spec-Driven: Providing a rigid technical specification for the agent to implement.
- Visual-First: Focusing primarily on aesthetics, animations, and visual polish.
The experiments primarily focus on modern frontend web technologies, utilizing frameworks and build tools such as:
- React / Vite
- TypeScript / JavaScript
- Vanilla CSS / CSS Modules
- Motion.dev (via MCP)
To view the final reports, metrics (motion-dev-mcp-benchmark.csv), and the interactive applications themselves, please check out the gh-pages branch. The static site hosted from there provides an executive summary and detailed performance comparisons across all agents.
This project is licensed under the MIT License - see the LICENSE file for details.