Private by default with hosted models

Choose how your project uses hosted models.

Velora can mask detected sensitive values in text before sending it to a hosted provider. Detection is not exhaustive, so review content before sending it. Choose a setting in Intelligence → Controls → Privacy with hosted models.

• Local only. Velora blocks hosted models for this project; local models can still work on it.
• Shielded (the default). Detected credentials, personal data and private terms in text become placeholders.
• Standard. Hosted models receive text as written, with recognised credentials replaced where detected.

How the shield works

1. Before text leaves your computer, Velora finds sensitive spans: recognised credential formats, values assigned to names like password or api_key, card numbers that pass the Luhn check, IBANs that pass their checksum, and the private terms you list for the project — client names, codenames, people.
2. Each becomes a placeholder such as [EMAIL_3fa9c21b], computed with HMAC-SHA256 from a key created on your computer that never leaves it. The same value is the same placeholder throughout a project, so the model can still follow who is who; the same value in another project gets an unrelated placeholder.
3. Replies come back with placeholders. Velora restores the real values on your computer — in streamed text, in summaries, and in the arguments of tools it runs for the model — so files it writes contain the real values. Tool results go back to the model shielded.

Use Check what a hosted model would see in Controls to try any text; it is computed on your computer and sent nowhere. Controls also list each connected model with what it may see, and how many values have been kept back so far — counts only, never the values.

Claude Code, Codex and other coding agents

Coding agents read files themselves, so in a shielded project they work in a shielded copy of it: placeholders in its files and paths, and a fresh Git repository without the project’s history, where old secrets would otherwise live on. When the agent finishes a turn, the files it created, changed or deleted come back into your project with the real values. If a file also changed in your project meanwhile, Velora keeps your version and saves the agent’s beside it as .velora-agent, with a warning in the task. Velora’s own tools are shielded the same way when an agent calls them. Turn this off in Controls with Shield project files from coding agents.

What the shield does not cover

Detection can miss sensitive information. Images and other binary files pass through unchanged, and text files over 2 MB stop a shielded copy from opening. Names require private terms. A coding agent may read outside the copy if its own sandbox permits it; linked dependency folders point into the project. Local only blocks hosted models in Velora, not network access by other tools.

Your data and intelligence · Agents and models · Download Velora