Chat app
A model answers.
PromptAnswer
user asksanswer returns
You get“Here’s a summary…”End of turn
TritonGPT or a browser chat tool is a conversation surface: prompt in, answer back.
Agent control surface
Work directly with your code, files, and systems
TritonAI Harness is UC San Diego's primary supported agent workspace. It runs on your computer and connects to campus models, local files, and approved campus tools under your supervision. It is in pilot on Mac (Apple Silicon) and Windows, and anyone with a Gateway key can request it.
Use it to work through a task in your project folder: read source material, prepare a draft, and check the result. You supervise the work and review consequential actions.
Prefer another client? Claude Code, Codex, Hermes, OpenCode, and other compatible clients connect to the same Gateway with the same key. Compare the client options.
You’ll need an approved TritonAI Gateway key to connect the Harness to campus models.
Already have a key? Download and set up.
Governed agent execution
- 01Local workspaceFiles and task history are stored locally; selected context is sent to the model.
- 02Campus gatewayZero-retention UC-hosted models on UC San Diego infrastructure.
- 03Task access modeChoose when the agent pauses for your approval.
Open the course
Self-paced training
Learn to use TritonAI Harness
Follow narrated lessons to run your first task, guide the work, and check the result. Then explore campus plugins, reusable skills, and app building.
Start with Essentials, then try the exercises using downloadable fictional files. Knowledge checks help you review what you’ve learned.
- Essentials
- About 8 minutes
- Full presentation
- About 24 minutes
Allow additional time for practice.
Stable release
Harness 0.3.7 stable release
Current stable version: 0.3.7. Release information checked October 15, 2026 against the published release. Available for macOS Apple Silicon and Windows x64. Nightly builds use a separate update track.
- The Tableau connection is no longer included in the stable app; it is only in the separate Nightly version. If you relied on Tableau, you will need the Nightly app to keep using it.
- This release adds a new Quiet Workspace theme to the app's appearance options.
- The saved list of skills now appears right away when you open it, so you spend less time waiting to pick one.
- When the agent asks for your approval, you now see the details of what it plans to do, helping you decide with better information.
- Sending feedback now goes through the TritonAI feedback skill, so your comments reach the team through a standard route.
To update an existing installation, open Settings > About > Check for Updates. For first-time setup, use the stable TritonAI Installer.
Computer use starts off. Enable it in Settings > General; on macOS, grant Accessibility and Screen Recording permissions and restart the Harness when prompted.
Read the Harness 0.3.7 release notes · Review approval modes
See the Harness in action
This narrated video runs a little over two minutes and shows results from the Harness app. It was recorded with fictional workshop files.
Try these tasks yourself in the full Harness training.
Tap Play to start. If the embedded player does not work, open the video directly.
The video starts muted with captions. Turn on sound to hear the narration. It shows a survey summary, an Excel dashboard, a briefing deck on the UC San Diego template, the five campus plugins, an n8n workflow explained step by step, a GitHub code review, and a sign-up app built on the campus design system and wired to an n8n workflow.
Introduction (0:00)
TritonAI Harness is UC San Diego's AI workspace on your desktop. You hand it real work, and it delivers finished results, using your files and campus systems.
Summarize a survey (0:11)
Here it is, live, with a made-up project: Jordan's fall workshop pilot. Qualtrics already reports the survey, but Jordan wants the team's own questions answered, so I hand it the raw results. Fourteen seconds later: averages for every question, the top comments, the missing data flagged, and a takeaway for the planning team, saved as a file.
Build an Excel dashboard (0:35)
Department leaders want one page, so: a presentation-ready Excel dashboard. It builds the workbook, and a few quick follow-ups polish the layout. Headline numbers up top, three charts in campus colors, and a one-line insight under each.
Create a briefing deck (0:52)
Jordan has ten minutes with leaders next week, so the meeting notes and the survey become a briefing deck on the official UC San Diego template. Nine slides with live charts, an owners table, a timeline to the next session, and speaker notes, all built from the source files.
Work with campus systems (1:10)
It plugs into the campus systems you already use: Microsoft 365, Google Workspace, Kuali Build, GitHub, and n8n. Here it reads an n8n workflow, explains each step in plain language, and confirms the last run succeeded. With GitHub, it reviews a code change: what changed, which files, and whether the build passed.
Build a web app (1:39)
And it builds software. A Microsoft Form would cover a basic sign-up, but Jordan's team wants the campus look and every sign-up in its own n8n workflow. Here's that app, on the official campus design system. Coming soon: self-service hosting, so the apps you build in the Harness can go live for campus.
Campus tools first (2:01)
When a campus tool does the job, use it. When you need it your way, build it with the Harness. That's TritonAI Harness: real work, finished on your desktop, with campus tools built in.
What a harness adds
The harness is everything around the model
The same approved models sit inside TritonGPT and TritonAI Harness. A chat app answers a prompt. The Harness wraps the model with context, tools, permissions, and memory, and runs a loop so it can act on real work.
Diagram key: The circled M represents the AI model.
Harness
A model works.
ContextToolsLoop
- files
- rules
- memory
- read
- edit
- verify
You getcontract.docxrouted for sign-offlogged
TritonAI Harness takes the same model and gives it a workspace, bounded tools, and a person who chooses the approval mode and reviews the result.
How the Harness works
These capabilities connect your request to the work in your workspace.
Diagram key: The circled M represents the AI model.
Loop
How does it keep going?
Takes a step, checks the result, and adjusts within the task’s permissions and stopping conditions.
Context
What does it work from?
Brings relevant files, instructions, and earlier results into the task.
Action
What can it access?
Uses enabled tools to work with files and connected services within granted permissions.
Scale
How are larger tasks organized?
Can divide planning, building, and checking among focused sub-agents.
Choose by task
Comparing TritonGPT and TritonAI Harness
Both services use UC San Diego AI infrastructure, but they serve different tasks and handle data differently.
| Capabilities |
TritonGPT
Web Platform
|
TritonAI Harness
Agent Workspace
|
|---|---|---|
| Where it runs |
Campus web servers
Access directly in any browser. No local installation or setup required.
|
Local workstation
Native desktop app installed on your Mac (M1+) or Windows x64 computer.
|
| System access |
Session uploads
Reads documents and spreadsheets uploaded into active chat conversations.
|
Local folders, terminal, & git
Reads project directories, edits local code, and runs shell commands according to the chat’s access mode.
|
| Data storage |
Campus server retention (90 days)
Chat conversations reside on campus servers and purge automatically after 90 days.
|
Local files and task history
Project files and saved task history stay on your device. Prompts and selected context are sent to the chosen model; connected tools exchange data with their services.
|
| Host plugins |
Web widgets & directory
Connects to university websites, departmental knowledge bases, and directory lookups.
|
Campus service plugins
Connect the included plugins listed below. Access follows your account permissions and enabled abilities.
|
| Human oversight |
Chat window interaction
You evaluate text responses directly in the browser interface.
|
Per-chat approval modes
Supervised pauses for file changes and write tools. Full access allows actions without approval prompts. Plugin abilities and account permissions still apply.
|
| Data classification |
P1 through P3 Approved
Approved for institutional campus records. Protection Level 4 (P4) is prohibited.
|
P1 through P3 Approved
Approved for administrative and research work in approved setups. P4 prohibited.
|
| Open TritonGPT Web | Get started with TritonAI Harness |
Campus roles
Who uses TritonAI Harness
The Harness provides structured workspaces tailored to different roles across UC San Diego.
Staff and citizen developers
Analyze spreadsheets, generate draft reports, and organize intake data using your files and enabled campus plugins. Choose a model route approved for the data you use.
Software developers
Build and host applications. Work across a repository with multi-file edits, test runs, and pull-request review, then move the result onto a supported campus hosting path. Start from the Skills Library and the development patterns TritonAI supports.
Faculty and researchers
Work with local datasets and run prompt pipelines through approved campus model routes. Check the requirements of your research project before using its data.
Core capabilities
Execution and safety architecture
TritonAI Harness balances agent capabilities with rigorous institutional guardrails.
Local-first execution
Project files and saved task history are stored on your workstation. The selected model receives the prompt and context needed for the task, including relevant file content or images.
UC-hosted models
The stable release uses GLM 5.3 Flash as the managed default and also offers GLM 5.3 and Glimmer. Existing selections of available models are preserved. Check the current models and routes for hosting and data-use guidance.
Supervised execution
Start with Supervised while learning: safe reads may proceed, while file changes, write tools, and commands needing additional permissions pause for approval. Choose the mode in each chat; exact behavior depends on the provider. Full access allows actions without prompts.
Credential isolation
The desktop app encrypts stored API keys and OAuth tokens with a key protected by your operating system. Connected tools receive authorized operation results; stored credentials are kept out of model context.
Host integrations
Included plugins
In Settings > Plugins, enable an included plugin, connect your account, and select its abilities under Access where available. Some abilities are granted during sign-in. Account permissions, enabled abilities, and the chat’s approval mode each control what the agent can do.
GitHub
Connect a GitHub account with OAuth to read repositories, draft pull requests, and automate development workflows. Access follows your signed-in permissions.
Google Workspace
Read UC San Diego Drive files, including Docs, Sheets, and Slides. Gmail and Google Calendar access are also available for accounts that use them. File access is read-only; unsent Gmail drafts and calendar changes are separate opt-in abilities.
Microsoft 365
Read Outlook mail and calendars with your signed-in permissions. Unsent email drafts, mail organization, calendar changes, and Teams chat access are separate opt-in abilities. The plugin has no email-send tool; sending to an existing Teams chat requires its own enabled ability.
UC San Diego Kuali Build
Connect with your API key to inspect apps, forms, documents, and workflow metadata. Creating, editing, and submitting Build documents require the corresponding write abilities.
n8n
Sign in through your browser to inspect and design workflows in UC San Diego’s n8n service. Running, creating, updating, and publishing workflows depend on the abilities granted during sign-in and the chat’s approval mode.
Reusable capabilities
UCSD Skills Library
Skills package instructions, references, and scripts that agents load on demand. Browse AI Team and Community skills in Settings > Skills. Skills you create or add locally stay under Your Skills; use Share with UCSD to submit a public contribution for review.
Accessibility compliance
Audit and remediate web pages, templates, and documents against WCAG 2.1 AA and UCSD Digital Accessibility standards.
Data classification
Classify application schemas, storage repositories, and API endpoints under UC IS-3 Protection Levels (P1 through P4).
Decorator branding
Build web interfaces matching UC San Diego Decorator 5 standards without altering protected campus chrome.
Automated code review
Run defect-first automated reviews before commits or pull requests merge into campus repositories.
Scale
Bigger tasks split across sub-agents
Planning, drafting, and verification can run as separate sub-agents, each with only the tools its job needs, and then roll back up into one coherent result. The person running the Harness still approves what leaves the workspace.
- PlanRead-only tools for reconnaissance: read, search, and browse approved sources. Nothing changes.
- BuildBounded execution for the everyday work: edit files and run commands inside the workspace you opened.
- VerifyExecute and read to test the result and confirm it before a person signs off.
Data governance
Privacy and data protection
TritonAI Harness is approved for UC Protection Levels 1 through 3 (P1 through P3) within approved campus setups. Protection Level 4 (P4) data is prohibited. Because your workspace files and transcripts reside on your local workstation, endpoint encryption and device security standards apply.