Rams MCP
Rams is a design reviewer for UI code. The MCP puts it inside your coding agent: the agent writes the code, Rams reviews it for design quality, scores it 0 to 100, and hands back fixes, all before a commit.
The Rams engine.
194 rules, 8 categories.
Not linter rules. Review judgments, versioned and public in the rules. New ones ship every week. Your agent gets sharper without changing a thing.
Semantic HTML, keyboard nav, focus, contrast
Token bypasses, contrast ratios, semantic use
Hierarchy, scale, line height, truncation
Scale adherence, gap usage, concentric radii
Composition, reuse, token leaks, prop APIs
Loading, empty, and error states, form patterns
Duration, easing, reduced-motion, layout shift
AI antipatterns: magic numbers, slop gradients
How it works.
Three moves, all inside the conversation you already have open.
One tool, review_files. The agent passes the UI files it just wrote or changed, straight from the working tree. Nothing to stage, nothing to push.
All 194 versioned rules across 8 categories, with server-side verification on the findings. The same engine, the same judgment, as every pull request review.
A real 0–100 score, findings with severity and file:line, and a concrete fix for each. The agent applies them and commits clean.
Works in every agent.
Anything that speaks MCP over Streamable HTTP. These are the ones we verify against.
Claude Code
Cursor
Codex
Windsurf
Zed
GitHub CopilotMCP vs the free Skill.
Rams also ships a free Skill: a design checklist your agent reads as it works. Both live in your agent. Only the MCP runs the real engine. Full comparison: Rams vs design skills.
Features.
Everything a review returns, and everything it never asks of your setup.
0–100, severity-weighted. Any critical issue caps it at 59, so 60+ always means zero criticals.
Severity, category, and file:line for every issue, plus a concrete fix ready to apply.
The score your agent sees locally is the score your pull request would get. No drift between local and CI.
Issues get fixed before they exist in history, not flagged after they ship.
An API key and one config block. No workflow files, no pipeline changes.
MCP reviews and pull request reviews draw from the same monthly pool. One plan, no surprise line items.
Set up in two minutes.
Create a key, add the server to your agent, and ask for a review. No CI config, no workflow files.
claude mcp add --transport http rams https://worker.rams.ai/mcp \
--header "Authorization: Bearer rams_YOUR_KEY"{
"mcpServers": {
"rams": {
"url": "https://worker.rams.ai/mcp",
"headers": { "Authorization": "Bearer rams_YOUR_KEY" }
}
}
}[mcp_servers.rams]
url = "https://worker.rams.ai/mcp"
http_headers = { "Authorization" = "Bearer rams_YOUR_KEY" }Any other MCP-capable client: Streamable HTTP at https://worker.rams.ai/mcp with your key as a Bearer Authorization header.
No commands to remember. Drop one rule into your agent's instructions and it reviews its own work: write, review, fix, re-review, commit.
Before committing UI changes, run a Rams review and fix every critical.● review_files · 3 files Rams score: 54/100 · 2 critical, 1 serious [critical/a11y] Banner.tsx:3 Clickable div blocks keyboard users [critical/color] StatCard.tsx:12 #22c55e bypasses your token layer [serious/ux] Hero.tsx:8 Two filled CTAs split the primary action ● Fixing all three, then re-reviewing… Rams score: 91/100 · 0 critical
rams review my changesCommon questions.
How the MCP relates to the Skill and the GitHub App, what it costs, and what leaves your machine.
The Skill is a free taste of the engine: surface-level, stateless, no score. The MCP is the full engine as a tool: every call runs all 194 versioned rules with server-side verification and returns a real 0–100 score, the same review a pull request gets.
No. They cover different moments. The MCP reviews code on demand inside your agent, before it is committed. The GitHub App reviews every pull request automatically, keeps score history, and covers your whole team. Most people who use the MCP also want the App as the backstop.
Each MCP review counts toward the same monthly review quota as your pull request reviews: free 30, Solo 50, Basic 300, Pro 750 per month. There is no separate MCP charge.
The files you (or your agent) pass to the review_files tool are sent to the Rams worker over HTTPS, reviewed, and returned with findings. Reviews are metered against your workspace. File contents are not used to train models.
Put the engine in your agent.
Sign up free, create a key, and your agent reviews its own work: 30 reviews a month on the free plan.
Frameworks