Security architecture
Useful access.
Explicit boundaries.
Donatello is an agent with tools, so its permissions must be visible, limited and reversible.
Project scope and permissions
Work starts in a selected project folder. Built-in file operations apply path checks; command execution and tools can require additional access. Review requests carefully: a project folder is not an operating-system sandbox for every program.
Protected credentials
Saved provider credentials use Windows-backed encryption. Do not place keys in project files or prompts. Provider tools and commands are not a guarantee that every third-party program redacts all sensitive output.
No inbound port
The Windows Agent makes outbound provider and account requests. Donatello never exposes a listening port on the user's router.
Revocable sessions
The Windows Agent uses short access tokens and a revocable device session tied to the Donatello account.
Visible permission decisions
When work needs access beyond ordinary project editing, Donatello asks the user to reject it, allow it for the current task or remember that type of permission. A task can always be stopped from the interface.
Local files are not the same as local inference
Project files are saved on your PC. Online AI services may receive prompts and relevant file or tool context. Local inference runs on your computer, but automatic mode is not a local-only privacy guarantee: it prioritizes available online routes. Review permissions and do not include secrets in prompts. Optional usage telemetry is disabled by default.
Donatello Local uses IBM Granite 4.1 3B Instruct, Q4_K_M, with a 16,384-token context. Activation downloads a model of 2,099,501,664 bytes (about 2.10 GB) and checks its SHA-256. The bundled runtime supports CPU and compatible GPU acceleration. Local inference can work offline after installation; online search, dependency downloads and cloud media still need Internet. This text-only model does not generate images, voice or video by itself.
Local previews may run a service on your computer. Do not expose a development preview to the Internet unintentionally. Review generated code before production deployment.
Setup and local AI routing · Current Agent conditions
Report a problem
Security reports can be sent to support@donatello.studio. Include the Agent version and a reproducible description; never include an API key.