Khimba/ trust

Inspect before trust

Know what the other model sees.

Khimba separates local preparation from remote inference. The confirmation is the boundary: before it, no source bundle is uploaded and no model charge is incurred.

Data flow

  1. The host agent resolves the requested reviewer, subject and focus.
  2. A local collector reads only the chosen paths. It respects ignore files and excludes common secrets, credentials, binaries and unsafe symlinks.
  3. The user receives a manifest of the proposed evidence and cost. Changing anything requires another confirmation.
  4. After confirmation, the encrypted request is sent to Khimba and the selected model provider for inference.
  5. The opinion returns to the host with usage and charge. Raw source bundles are deleted after the opinion runs.

What the reviewing model cannot do

The reviewer receives supplied text. It cannot browse arbitrary URLs, read additional local files, invoke the user’s tools or silently expand the scope. A model may infer sensitive facts from supplied content, so users should still inspect the manifest and avoid unnecessary sensitive data.

Pricing and model identity

Khimba reports the selected model’s current input and output rates during preparation. The user approves a maximum. The final charge is measured official-provider cost plus a 30% platform markup, and unused reserved credit is released. If a requested model is unavailable, Khimba reports that condition instead of silently substituting another.

Limitations

Filtering reduces accidental disclosure but cannot recognise every secret or sensitive business fact. Model output is AI-generated and can be incomplete or wrong. Users should reproduce technical findings and obtain qualified human advice for professional or high-stakes decisions.

Questions and reports

Read the full privacy notice and terms. Send security concerns or data questions to support@khimba.ai. Product feedback is submitted only when a user explicitly asks; it is never inferred from conversation.