Skip to main content

FAQs

Frequently Asked Questions

Who can request TritonAI Developer API access?

UC San Diego faculty, staff, researchers, and sponsored student projects may submit a request. Other UC campuses may request Gateway access through an inter-campus recharge agreement. CSU, affiliate, and partner requests may also be reviewed. Eligibility and the appropriate access path are confirmed from the purpose, funding, affiliation, data, owner, and intended audience. Medical Center clinical and patient-care operations use the Medical Center's platform.

How do students request Gateway access?

Students interested in campus-hosted models for a project should apply through RAIN (Real World AI Network). Sign in with your UC San Diego email and submit the Raindrop AI resource request. RAIN reviews student eligibility and the project, then issues approved keys through the portal. Access depends on pilot availability and the limits assigned to your key.

See the student access guidance on Get Started. For application questions or request status, contact Aryan Pershad at RAIN.

What data can I use?

The Gateway, TritonGPT, and TritonAI Harness are approved for UC information classified from Protection Level 1 through Protection Level 3 (P1 through P3) when the work is within an approved setup. P4 data is not approved. That approval covers the platform. The person or office responsible for the information still decides its classification and permitted use. Describe the data in your request so the service team can identify the appropriate hosting choice, controls, and any required privacy, security, or service-owner review. Restricted or health information requires a separately approved service path. Review the UC Protection Levels.

Do I have to use TritonAI Harness?

No. TritonAI Harness is the primary supported client, and Claude Code and Codex are supported alternatives. Hermes, OpenCode, and any other client that accepts a base URL and API key can connect to the Gateway with the key from your approval. Features, extensions, and setup differ by client, and support for those clients comes from their own documentation.

Where can I see what changed in Harness?

The Harness release notes summarize stable desktop releases and link to the original release details. Find the downloads and setup instructions on Get Started.

When does the Harness ask for approval?

Choose the mode per chat in the message composer. A chat created from within another chat inherits its mode. Other new chats use your last choice on that device or the app or project default. Without a saved override or remembered choice, the default is Auto. Full access allows enabled actions without approval prompts. Use Supervised while learning: safe reads may proceed, while file changes and write tools pause for approval. Command handling depends on the provider. Auto-accept edits permits file edits automatically; Auto handles routine actions automatically with provider-dependent review. Account permissions and plugin capabilities still limit which tools are available. Read the Harness permission mode documentation.

How do I enable computer use?

Computer use is optional and starts off. Open Settings > General and follow the readiness instructions. On macOS, grant Accessibility and Screen Recording permissions and restart when prompted. The stable release supports computer use with the Codex provider on a local desktop backend. If the setting says Needs permissions, complete the listed steps. Screenshots can be sent as model context; check the privacy statement.

Should I install stable or nightly Harness?

Use the stable TritonAI Installer for the campus setup path. The current stable Harness release is 0.4.2, with release information checked November 17, 2026. Nightly builds are prereleases with separate application data and update tracks. Update the Harness app through Settings > About > Check for Updates; app and engine updates are separate.

Why is a plugin connected but a tool unavailable?

In Settings > Plugins, check that the plugin is enabled and finish its sign-in or API-key setup. Review its readiness status and granted capabilities. Some capabilities require separate authorization, and your account must have access to the source. Read access does not grant write access. A task’s approval mode also controls whether an available action can proceed. Share the readiness message with tritonai@ucsd.edu if you need help; never include a secret key.

What's the difference between UC-hosted and enterprise cloud models?

Both may be available through the common gateway. They differ in capability, hosting, rate, and approved controls. Cloud and UC-hosted access are approved separately. Review the TritonAI model catalog for the current catalog and published rates, then confirm the approved access and limits during intake.

How do purpose and funding affect charges?

The intake classifies campus administration, instruction, sponsored research, non-sponsored research or faculty exploration, clinical or patient-care operations, and external partner work separately. For campus administrative work, UC-hosted token use is covered and not recharged. Cloud use is charged from the first token at the current market rate published in the Model Hub. Research projects charge both UC-hosted and cloud model use to a grant or approved research project chartstring. Other UC campuses may use the Gateway through an inter-campus recharge agreement and are recharged for both UC-hosted and cloud model use. Your approval response records the access, limits, and billing treatment assigned to your key.

What does a cloud access request require?

Cloud access requires a chartstring, named budget owner, and spend limit when the key is issued. Chargeable requests also require the appropriate budget-owner or principal-investigator acknowledgment. ITS cloud requests follow the TritonAI budget-manager approval path and use the designated program funding process.

How are Health Sciences requests classified?

Health Sciences teaching, research-support, and administrative requests are evaluated by purpose like their campus equivalents. Medical Center clinical and patient-care operations use the Medical Center's platform. Incidental TritonAI use for that work requires case-by-case review.

Can I use the API for a production application?

A working prototype does not by itself authorize production use. Production readiness is reviewed from the application's audience, data, integrations, accessibility, security, support model, monitoring, and accountable owners.

How long does review take?

Timing depends on the use case and any required project-specific review. Submit complete information about the task, data, audience, sponsorship, and ownership to help the team identify the next step.

How do I get help or report issues?

For support inquiries or to report issues, email tritonai@ucsd.edu to create a support ticket.

Submit an Access Request

Take the first step toward integrating AI into your work at UC San Diego.

Request API Access

Stay Informed

To receive the latest announcements and news, subscribe to the TritonAI mailing list.

Subscribe

Current stable release guidance

TritonAI Harness 0.4.2.

Read the release notes

  • Computer use is off by default; turn it on in the desktop app under Settings > General. It works only with the Codex provider and only on the local desktop backend, so WSL, SSH, and other remote backends do not receive it.
  • Permission modes are set for each chat from the message composer. New chats remember the last mode you chose on that device, and a chat created inside another chat keeps that chat's mode. Until you choose, new chats use your app or project default, which starts as Auto.