What Onefold reads, and what it stores

Onefold asks for access to the repositories that hold your product's interface. This page is the straight account of what happens to them, including the parts that are not finished yet.

What it reads#

• Repository contents, read only, on the repositories you select during GitHub App installation, through short lived installation tokens. No write access is requested and none is used.
• Figma libraries you connect: components, component sets, variants, variables and descriptions. Published libraries, not your working files.
• Storybook, discovered from the repository itself. There is no separate account to connect and no running instance to point at.

It does not read issues, pull requests, CI logs, secrets, your product's user data, or analytics.

What it stores#

Onefold stores the structured result of the analysis rather than a copy of your repository:

• component, token, story and documentation records, with the file path and line range each one came from,
• the relationships between them, including design to code mappings and their confidence values,
• the findings behind each readiness score,
• short excerpts of source, of the order of a few lines, where a finding needs to quote the evidence that produced it.

There is no full repository copy at rest. Analysis works against a shallow checkout that is discarded when the run ends.

What goes to a model provider#

Discovery is deterministic. Parsers walk the code and produce structured facts, and no part of that step involves a model.

Models are used for the narrow questions that need judgement: classifying what a component is for, proposing a mapping between a Figma component and an implementation, summarising a finding. Those calls carry the minimum context the specific question needs, which is typically a component name, its props, and a documentation excerpt, not files and not the repository.

Model output is never treated as source truth. Every assertion records whether it came from a source fact, a deterministic derivation, a model inference or a human confirmation, and human confirmation outranks model inference.

Tenant separation#

Every record belongs to a workspace, every workspace belongs to an organization, and every query is scoped by both. Authorization is enforced in the data layer rather than in the interface, so a missing check in a route cannot expose another tenant's graph.

Deletion#

Disconnecting a source deletes the observations and graph nodes derived from it. Deleting a workspace deletes everything in it. Both are immediate rather than scheduled, and neither leaves an orphaned analysis behind. During early access, write to onefold@pitsch.me and it is done the same day.

What Onefold does not have#

No SOC 2 report, no ISO 27001, no penetration test, no data processing agreement template yet. It is in development, and describing a compliance posture it does not have would be the least trustworthy thing on this page. The security page tracks what changes as it changes.