witnora

Private Capability Discovery

Formal privacy commitment

Witnora does not require source-code access for its default capability discovery. In private_discovery mode, discovery runs inside the customer environment and uploads only a signed metadata snapshot. Witnora does not upload source code, prompt text, credentials, tool inputs, tool outputs, database rows, or file contents through this path.

The uploaded snapshot is limited to:

This promise applies to the versioned witnora.private_capability_snapshot.v0.1 path. Other evidence paths carry their own declared privacy classification and limitations.

Customer choices

witnora onboard performs the first local scan automatically. The scan reads package/dependency metadata and top-level repository names locally, converts matches into generalized capability groups, and uploads only the signed snapshot described above. It does not traverse or upload source files.

Later refreshes are explicit:

npx witnora@latest discover

If a customer-owned Gateway is online, it may consume a short-lived Hosted refresh request. Otherwise, Hosted displays the command and the request expiry rather than pretending it can remotely inspect a laptop.

Runtime only

No repository scan is performed. Witnora learns capability metadata only when calls cross the Universal Coverage Gateway. This provides the strongest repository isolation, but coverage grows only as paths execute.

Source-assisted (optional)

Source-assisted discovery is disabled by default. It requires a separate, explicitly reviewed GitHub App authorization with visible repository scope and permissions. Witnora never silently upgrades private discovery into source access.

What discovery proves

A completed snapshot proves that a customer-held Ed25519 identity reported the listed metadata for one repository fingerprint and one short-lived challenge. It helps identify likely capability groups and unknown paths.

It does not prove that every runtime path was observed, that an action was mediated, that an outcome was independently verified, or that the agent is safe. Those claims require runtime Gateway evidence, outcome probes, and a reviewed assurance case.