I shipped a free MCP server for devs. It's called Rocky.
Rocky is a pre-PR quality gate for AI coding tools. Same model, about 65% fewer tokens on a measured feature. Free, open source, MIT. Works with Cursor, Claude Code, VS Code, and ChatGPT.
I shipped a free MCP server for devs. It's called Rocky.
The problem
AI coding tools are fast. Most of what they generate still gets rejected: wrong conventions, inconsistent patterns, missing tests, structure your reviewer will not accept.
You burn two things. Review time on bad output, and tokens regenerating it.
What Rocky does
Rocky is a quality gate that runs before a human sees the code. It does not write code. It checks what your AI agent produced against your stack standards, agents, and rules, before you open a PR.
Opinionated on purpose. An AI with no guardrails writes code for itself, not for your reviewer.
Why MCP matters for performance
Without it, the AI walks the whole repo: 10 to 18 file reads per session, full test suites every loop, terminal logs dumped into context.
With Rocky MCP you get structured tool responses instead of raw file reads, scoped tests instead of full suites, and a pre-PR gate with a clear ready / not-ready verdict.
Measured on a real production repo
- Discovery tokens: 125k to 14k (-89%)
- Validation time across 8 loops: 232s to 77s
- Total tokens per feature: about 255k to about 89k
- Sonnet cost per feature: about $1.62 to about $0.85
Same model. Same code. Roughly 65% fewer tokens.
What you get
- Pre-PR quality gate (ready / not-ready)
- 24 agents and 26 skills your AI can reference while generating
- One
devkitgateway instead of a pile of tools - Works with Cursor, Claude Code, VS Code, ChatGPT
Free. Open source. MIT.
- Repo: github.com/hellozheat/rocky
- Install and security: docs/INSTALL.md
- Full measured report: Rocky MCP value report