Trinity Hub¶
Trinity is a multi-tenant, multi-agent workspace where scientists and AI agents collaborate to run work on DOE leadership computing facilities. You describe a task in plain language — "run a Quantum ESPRESSO relaxation on Polaris", "build LAMMPS on Aurora", "is Frontier up?" — and Trinity plans it, submits and monitors the HPC jobs, moves the data, and reports back, all from a chat window in your browser.
Trinity Hub (this site) is the open front door to that platform. It documents how to use Trinity as a scientist, and how to extend it by publishing your own agents and reusable skills.
Open Trinity How to use Trinity
What Trinity is¶
Built at the Argonne Leadership Computing Facility, Trinity runs on the Claude Agent SDK for reasoning and tool use, and rides Globus for identity, data movement, and compute — so one authenticated conversation can reach every facility you're authorized to use.
It composes reusable building blocks rather than improvising from scratch each time:
- Agents — domain specialists (DFT, molecular dynamics, performance engineering, storage benchmarking, and more) that plan and run multi-step campaigns, alone or together with people in a shared room.
- Workflows — curated, parameterized HPC pipelines for common tasks (builds, submissions, ports) that agents draw on instead of writing everything from scratch.
- Skills — atomic, reusable capabilities — a procedure, a tool wrapper, a known fix — that any agent can load.
Every action is recorded as Knowledge (validated domain facts that improve future runs), Traces (a full timeline of tool calls and observations), and Provenance (why a decision was made) — so results stay auditable and reproducible, not just "the agent said so." See How it works for the registry-vs-runtime split that lets outside agents and skills plug in.
Pick your path¶
-
Use Trinity (scientists)
Log in with Globus, ask for compute work in natural language, and watch jobs run across ALCF, NERSC, and OLCF.
-
Extend Trinity (agent & skill authors)
Publish your own agent so others can
@mentionit in a room, or share a reusable skill that any Trinity agent can load. -
Operate Trinity (deployers)
Understand the registry-vs-runtime split, sync the registry into a running server, and review the trust boundaries.
-
See it in action (anyone)
Real DFT, molecular dynamics, GPU porting, and paper-reproduction campaigns run end-to-end through Trinity — the prompt, what happened, and the result.
What Trinity does¶
| You ask… | Trinity… |
|---|---|
| "Relax this structure with Quantum ESPRESSO on Polaris." | writes the input + PBS script, submits via the facility API, monitors to completion, returns results. |
| "Build NekRS on Aurora." | picks the right modules/toolchain and iterates the build on a compute node. |
| "Train GPT-2 on 4 Aurora nodes." | stages data, generates the distributed launch, submits, and tracks the run. |
| "Is Perlmutter up? What's the queue?" | queries facility status and reports availability and load. |
| "Move these outputs to my laptop." | runs a Globus transfer and reports the task status. |
Day to day, Trinity runs an embedded coding agent with tools for the ALCF/NERSC/OLCF integrated research-infrastructure APIs, Globus transfer & compute, and experiment tracking — but you never touch those directly. You stay in the chat.
Supported facilities¶
- ALCF — Polaris, Aurora, Crux, Sophia, Sirius
- NERSC — Perlmutter
- OLCF — Frontier
A platform built for many users¶
Trinity serves many scientists at once under strict tenant isolation: every user has private sessions, files, credentials, and execution traces, while sharing a read-only knowledge base and a library of reusable workflows. Access is by Globus single sign-on — your existing DOE facility identity.
New here?
Start with Getting started to log in, then skim The chat interface to learn how to ask Trinity for work.
Maintainer¶
Trinity and Trinity Hub are created and maintained by Huihuo Zheng (@zhenghh04), Argonne Leadership Computing Facility (ALCF) / Argonne National Laboratory. See GOVERNANCE.md for how contributions are reviewed, and CONTRIBUTING.md to get involved.