[WIP] feat: mcp server prototype

[WebEferen]

This PR introduces @ripple-ts/mcp-server, a Model Context Protocol (MCP) server that exposes Ripple's compiler capabilities to the LLMs

What it brings:

• Compiler Integration: Allows AI assistants to compile Ripple code to JavaScript/CSS directly through MCP tools, giving them the compilation results they understand with fewer tokens.
• AST Parsing: Provides access to Ripple's parser for syntax analysis and code understanding.
• Component Generation: Enables AI-assisted creation of new Ripple components with proper file structure.
• Reactivity Analysis: Analyzes Ripple code to identify reactive variables, tracked collections, and potential issues.
• Documentation: Allows AI assistants to reach out to the current documentation for up-to-date details.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Preview	Comments	Updated (UTC)
ripple	Ready	Preview	Comment	Nov 25, 2025 2:55pm

[jsudelko]

My intuition tells me a user-editable AGENTS.md and functional LSP linting gives a lot more value than an MCP server. Ripple-bench?

Consensus is brewing among LLM user/providers that MCP was brilliant on paper, but it makes bots less intelligent and more costly in practice.

For example, I'd wager none of the currently included tools improves performance to justify including them in all LLM requests.

[WebEferen]

@jsudelko MCP server would be very beneficial for the fast-changing early-stage framework as we are not gonna be well recognized by the LLMs (yet). This will add additional layer of knowledge for the LLM to work without polluting AGENTS.md / llms.txt files (as they need to stay as clean as possible).

Here the role of the MCP is to perform some of the actions and checks for the user. By giving it a knowledge base, we are making sure that our rapid schema changing will not affect the agentic coding capabilities.

We can argue about how effective or ineffective the MCP servers are, but that's not the point. We need a tool that would give the users quicker and faster way to jump/propose/orchestrate their work and reduce initial hassle with some of our own internal choices (for example "@" [reactivity], different mental model, compat layers etc.).

This server will have more features (such as auto-refactor for react-like modules/components that plays well with LLMs like Claude / GPT / Gemini and allow users faster adoption for their codebases).

For now it's size it's not that bloated, as the complete tools definition is roughly ~316 tokens in size.

Another benefit is that you as an user, have the choice to either connect the MCP server or use pure agentic workflow with AGENTS.md / TASKS.md files.