Roast is a design evaluation agent. Point it at a live URL, an image, a Figma frame, or a design-token file, and it captures the actual interface, screenshots it, and runs it through a visual evaluation engine. You get back structured critique across six dimensions — typography, color, spacing, components, hierarchy, and anti-patterns — with severity ratings and concrete fixes. It is not a set of rules you bolt onto a prompt; it is a system that looks at what you built and tells you, and your agent, what to change.

Roast connects through MCP (Model Context Protocol), so it plugs into Cursor, Claude Code, Windsurf, or any MCP-compatible client with one JSON block. Once connected, your agent can call it on its own: roast_url, roast_image, roast_figma, and roast_tokens each run an evaluation. For bigger jobs, roast_start_session opens a review loop that captures your v1, checks it against your stated goals, and keeps scoring every iteration until the design meets them — then roast_close_session writes the final report. The agent does not need to be a designer; it just needs to read a structured result and apply it.

It evaluates anything visual it can see: rendered URLs, screenshots, Figma frames, and token files. The engine scores typography, color and contrast, spacing and rhythm, component consistency, visual hierarchy, and common anti-patterns, then checks the result against a corpus of 50,000+ validated UI patterns so every verdict is anchored to real references. It does not write your code, and it does not enforce brand opinion — it flags what is objectively weak and returns the exact replacement so your agent can decide how to apply it. Context matters: pass a short context string (e.g. "fintech dashboard") and the evaluation calibrates to the right expectations for that kind of product.

One endpoint, three shapes of output. Mode "Diff" returns element-level fixes with exact replacements — the fastest thing to hand an agent. Mode "Score" returns a numeric breakdown across dimensions so you can see where the design stands and what is dragging it down. Mode "Profile" returns a full read of the design's aesthetic and identity and where it should go next. Same evaluation engine underneath; you pick how much detail you need. If you are not sure, start with diff — it is the default and the most actionable.

It is trained on a structured corpus of 50,000+ design references from some of the most precise interfaces in production. Every evaluation returns the closest corpus match as a reference point, so a verdict is never just an opinion — it is an evidence-backed comparison against real design work. That is the difference between a taste prompt that regurgitates rules and a system that measures your actual output against a reference set.

Evaluations run on a simple credit model: 1 credit per roast_url, roast_image, roast_figma, or roast_tokens call, and 2 credits to open or iterate a session. Searching the design corpus and getting brand guidelines are free. Credits are tracked against your API key and you can top up from the dashboard at designroast.cc/billing. Compared to a human design review that takes days, a full evaluation costs pennies and returns in seconds.

Roast sees only what you explicitly evaluate. For a URL eval it captures a screenshot of the page; for an image or Figma eval it sees that image or frame. It does not read your repository, your prompts, or anything you have not asked it to look at. The design agent evaluates the artifact you hand it, not your codebase.