Onefold
Early access

How it works

How Onefold works

Five steps between connecting a repository and an agent asking your design system which component to use. Nothing is migrated, and nothing is re-authored.

The run, end to end

  1. 01

    Connect the sources

    Install the GitHub App, pick the repositories that hold your system, and optionally connect the Figma libraries that describe it. Selection is per repository, permissions are read-only, and nothing is analysed until you say so.

  2. 02

    Discover what is there

    A background run walks the repository and parses it: components and their props, variants, tokens and theme files, Storybook stories, READMEs, MDX and any agent instructions already committed. Parsers do this work, not a model, so the result is the same every time.

    Components84 found
    Buttonexported
    DateRangePickerexported
    Modalno docs
    StatusPillnot found

  3. 03

    Reconcile design with code

    Figma components, component sets, variants and variables are matched against the implementations that realise them. Names, variant structure and descriptions produce candidate mappings, each with a confidence value and the evidence behind it.

    MappingFigma to code
    ButtonButtonconfirmed
    DialogModallikely
    Tooltipno matchunmapped

  4. 04

    Score, and show the evidence

    The eight readiness dimensions are calculated from what was found. Every number decomposes into findings, and every finding links to the file, story or Figma node that produced it.

    Evidence98% confidence
    ButtonFigma / Core / Button
    • Identical component name
    • 5 of 5 variants match
    • Description describes the same purpose
    • Confirmed by a maintainer

  5. 05

    Hand it to the agents

    The parts of your system that are ready become retrieval-oriented context, exposed over MCP so Claude Code, Codex and Cursor can query it directly instead of guessing from a component name.

    search_componentsMCP

    intent: “let someone pick a date range”

    DateRangePicker0.97

    avoid DatePicker

    DatePicker selects a single date only.

What a run produces

A component inventory
Every component Onefold could find, where it lives, what it exports, its props and variants, the stories that exercise it and the documentation that mentions it.
A system graph
The components, tokens, stories and documents as nodes, with typed relationships between them: implements, uses token, has variant, documented by, composed of.
A readiness assessment
Eight dimensions, each with its own evidence, combined into one score. The model is documented in full on AI Readiness.
A findings list
The specific gaps, ordered by how much they cost you: components with no usage guidance, no negative guidance, no examples, literal values instead of tokens, ambiguous neighbours.
A run record
What was analysed, when, how long it took, what was discovered and what failed. Analysis is observable rather than magic.

Parsers first, models second

The order matters more than it sounds. Discovery is deterministic: an AST walk finds exports, props and token references, and it finds the same ones tomorrow. Models are used where judgement is genuinely required, such as deciding whether two similarly named components mean the same thing, and they are given the minimum context that specific question needs.

This is the difference between a tool you can trust with a hundred-component system and a demo that works on ten. It also means your repository is not shipped wholesale to a model provider, which is covered in detail on Security.

Provenance is not optional. Every assertion records whether it came from a source fact, a deterministic derivation, a model inference or a human confirmation, along with its evidence and confidence. Ask why a mapping exists and you get a list you can check. A human confirmation always outranks a model inference.

What Onefold is not

  • Not a component library. It hosts nothing and becomes no part of your build.
  • Not a documentation site. Your docs stay where they are. Onefold reads them.
  • Not a Figma replacement. Design stays in Figma.
  • Not a chatbot on top of your design system. The intelligence sits in the product and in what agents can query, not in a chat window.

Run it against a real repository

Early access begins with GitHub. Connect the repository that holds your design system and read the first report.

Request early access