Data architecture
TritonAI Harness uses a local-first architecture. Local project files and task history are stored on your computer. Selected context is sent to the model, and enabled plugins communicate with their connected services. This statement explains what data stays local, what passes through the gateway, and how UC policies apply.
This statement describes local desktop use with approved campus model routes. Connected services retain their own records, and other providers or cloud tasks have separate data-handling terms.
Local storage
Local desktop task data is stored within your user account. This local storage is separate from context transmitted for model requests and records held by connected services.
Project files stay in their directories on your computer. The agent can read or change files through its tools according to the task approval mode and granted access. Files and attachments selected as context may be sent to the model.
Task transcripts, agent execution steps, and interaction histories are written directly to your local application directory. They are not stored on central campus servers.
Memory saves daily summaries as local Markdown notes. Relevant notes may be included in model requests. Settings > General > Memory controls collection and optional OneDrive sync. See the current stable release guidance below for defaults and changes.
Network transit
Communication between your workstation and TritonAI infrastructure is limited to active model requests and operational telemetry.
When an agent turn executes, selected context travels across encrypted HTTPS to the TritonAI Gateway. It can include conversation text, file contents, and tool results. Image attachments and computer-use screenshots can also be sent to models that support them.
UC-hosted models on campus infrastructure process requests with zero data retention and no model training. Your prompts are discarded after generating the response. For UC-hosted open-weight routes, inference, request context, and responses remain within the UC-managed environment and are not sent to the model provider or any third party, including providers outside the United States. UC-hosted weights are run locally, not accessed as an external hosted service. Other approved open-weight routes may use enterprise cloud providers under UC agreements.
The Harness maintains an empty cloud fallback list by default. If a UC-hosted model is temporarily unavailable, the error is surfaced immediately. Private campus context is never silently redirected to external cloud providers.
The Gateway logs request timestamps, model identifiers, token counts, and account identifiers to monitor platform availability, manage quotas, and apply recharge billing where applicable.
The Gateway does not record the text content of your local files, shell command outputs, or conversation transcripts in its telemetry logs.
Credential boundaries
TritonAI Harness uses host-managed authentication to isolate credentials from language models.
The Harness stores managed plugin credentials in encrypted local files, with the encryption key protected by the operating system. Its credential system supplies secrets to authorized operations and keeps stored credential values out of model context.
The Harness application process makes authorized API calls directly. The AI agent only proposes structured parameters for each action.
Email draft tools create unsent drafts. Calendar write tools make real changes when enabled. Available actions depend on account permissions and plugin capabilities; the task approval mode determines when the agent asks. Choose Supervised for write tools to pause for approval. Full access allows enabled actions without approval prompts.
Each plugin can be toggled on or off in Settings. Disabling access immediately revokes tools before the next agent turn.
Policy compliance
Use of the Harness must align with University of California Electronic Information Security Policy (IS-3).
TritonAI Harness is approved for UC information classified as Protection Level 1 (Public), Protection Level 2 (Internal), and Protection Level 3 (Sensitive) within approved campus setups.
Because project files and transcripts reside on your local computer, your workstation must meet UC San Diego minimum security standards, including active disk encryption (FileVault on macOS, BitLocker on Windows).
Protection Level 4 (P4) data is strictly prohibited in TritonAI Harness. This includes electronic protected health information subject to HIPAA, credit card numbers subject to PCI DSS, and high-consequence regulatory data.
For health system patient-care workflows, consult the approved clinical AI services available through the UC San Diego Health Pulse portal.
User control
Manage local task history, project files, and connected-service records separately.
Deleting a project folder does not establish that all associated task history, attachments, or application caches have been removed. Backups and records in connected services may retain separate copies. Follow your department’s retention requirements, and contact the service team if you need help removing saved data.
To revoke Microsoft 365 access, use the Disconnect button in Settings under Plugins, or revoke permissions directly in your Microsoft 365 account security settings.
Governance contact
For questions regarding TritonAI data governance, security assessments, or approved use cases, contact the service team.
TritonAI Harness 0.3.7.