What it is
designmd.supply is an MIT-licensed Next.js app that turns a public domain into a DESIGN.md file. Its README describes the result as a Google-spec document ready to feed to an artificial intelligence (AI) agent. It is best understood as a reference implementation for Context.dev’s web-data application programming interface (API), not as a separate commercial product.
The live site was not reachable during the research. Direct requests hit a Vercel-style rate-limit checkpoint (status 429), and the reader proxy hit a Cloudflare challenge. This review therefore relies on the public repository, its source code, and Context.dev’s own material.
First principles
A generated DESIGN.md is made of three inputs and one step. The inputs are evidence about a site, a prompt, and a model. Here the evidence is Context.dev’s extracted styleguide and brand data, a screenshot, and the page as Markdown. The step is the model writing a document from that evidence.
That structure sets the limits. The output can be exact only where the evidence is exact. Where evidence is missing, the model has to infer, and an inferred value in a token field looks identical to a measured one. designmd.supply’s prompt asks the model to prefer extracted values and to state uncertainty in prose, not in token values. The tokens stay clean, but only the prose says which of them were inferred.
Every uncached domain also costs upstream calls: four Context.dev APIs and a model call. That is why the pipeline caches results, and why code that is free still needs metered keys to run.
Finally, a generator is only useful to an agent if the agent can reach it. A site that blocks automated access, with no command-line or tool interface, leaves the agent nothing to read.
Who is behind it
The repository belongs to Context.dev, and its license says Copyright 2026 Context.dev. As of September 2026 it had 298 stars and 31 forks. It was created on 2026-05-05 and last pushed on 2026-09-13. Context.dev describes itself as a unified web-context API for software and AI agents. The research identifies it as Y Combinator-backed.
Third-party search snippets name Yahia Bakour as founder. That attribution is UNVERIFIED from a primary Context.dev source. The project itself offers the more useful fact: its source code, prompt, dependencies, and architecture are open to inspection.
Format and how an agent consumes it
The literal generation specification lives in lib/design-md.ts. It describes DESIGN.md this way:
export const DESIGN_MD_SPEC_SUMMARY = `DESIGN.md is a self-contained plain-text representation of a design system. It contains optional YAML frontmatter with normative machine-readable tokens and a Markdown body with human-readable rationale.
YAML frontmatter:
- Must begin and end with a line containing exactly ---
- Common top-level keys: version, name, description, colors, typography, rounded, spacing, components
- Color values must be SRGB hex strings beginning with #It then directs the Markdown body to use these sections: Overview, Colors, Typography, Layout, Elevation & Depth, Shapes, Components, and Do’s and Don’ts. The prompt names a Context.dev extracted styleguide as primary evidence, with a screenshot and the page’s Markdown as supporting evidence:
Requirements:
- Return only DESIGN.md content, no commentary before or after it.
- Include YAML frontmatter with version: alpha, name, description, colors, typography, rounded, spacing, and components.
- Prefer precise values present in the Context.dev payload.
- If data is missing, infer conservatively and state uncertainty in prose, not in token values.The documented pipeline is plain. A domain is checked against a Turso cache. On a miss, the Styleguide, Brand, Screenshot, and Markdown APIs run in parallel. The results are assembled into a prompt, sent through Vercel AI Gateway, written as DESIGN.md, and cached.
designmd.supply gives an agent no dedicated command-line interface (CLI) or Model Context Protocol (MCP) server. An agent can use the generated Markdown from the web page, or a team can self-host the app from its repository. Context.dev separately offers software development kits (SDKs) and an MCP integration for raw web data, but that is a Context.dev product, not a designmd.supply interface.
Pricing and access
The app itself is free, MIT licensed, and self-hostable with npm install && npm run dev. Running it needs a CONTEXT_DEV_API_KEY, a VERCEL_AI_GATEWAY_API_KEY, and optionally Turso credentials. Context.dev’s pricing was: Free at $0/month for 1,000 credits/month; Developer at $25/month for 10,000 credits; Pro at $149/month for 200,000 credits; Scale at $499/month for 1,000,000 credits; and Enterprise custom. A base scrape costs 1 credit and browser actions cost 2.
What works
The main strength is transparency. The generation prompt and the document structure are open, not hidden behind a hosted interface. The implementation also states an important rule about evidence: prefer precise extracted values, and state uncertainty in prose when data is missing.
The architecture is reproducible. A team can inspect the cache, the source APIs, the model boundary, and the environment contract instead of trusting a black-box extraction. Self-hosting helps teams that need control of the pipeline.
What does not
The flagship hosted site blocked automated access during this research. That weakens the case for a generator meant to serve agents. The app has no first-party CLI or MCP server. Generation for arbitrary domains beyond the visible demo and cache is UNVERIFIED, as is a downloadable result from the unavailable user interface.
“Free” also describes the code, not necessarily the running system. Running it depends on a metered upstream API once the free allowance is used. Its maintenance follows Context.dev’s priorities.
What we take from it
Velofy could propose an open reference implementation with a documented pipeline, environment contract, and literal prompt. That would make a design-spec proposal inspectable and forkable. It could also propose a clear separation between observed tokens and conservative inference.
The guardrail is operational: material meant for agents must stay reachable by agents. Any future Velofy design extractor should publish stable files and a direct programmatic path. Velofy has not shipped such an extractor, package, or MCP server.