November 5, 2026 · 3.0.0-beta.16

Private by default, even with Claude and ChatGPT.

Each project decides what hosted models may see, and the models you train are scored, exportable and ready to train again.

• New

  Choose Local only, Shielded or Standard for each project. Detected sensitive values in text can be replaced with placeholders and restored on your computer.

• New

  List private terms — client names, codenames, people — and preview how sample text is masked.

• New

  Claude Code, Codex and other coding agents work in a shielded copy of a shielded project, and their changes come back with the real values.

• New

  Sync an internal wiki, documentation server or JSON endpoint into Sources, with its token kept in your operating system’s keychain.

• New

  Every training run is scored against its base model on your own examples, held-out examples first, so you can see whether training helped.

• New

  Export a trained model as a standalone model with a model card, then serve it to your agents or train it again on new data.

• New

  Carry an approved dataset to another base model and compare the two scorecards.

• Improved

  Draft training examples from a source instead of writing each one, and train from a form that opens on your newest model and dataset.

• Improved

  Intelligence opens on a four-step path: connect your data, give your agents context, train a model you own, and keep it private.

• Fixed

  Models served on your own computer, including ones you trained, now count as local for privacy and routing.