AI agents and integrations

From answering questions to submitting jobs.

We build AI assistants and agents for research computing programs. They answer from your documentation, read live data from your systems, and act for a user with that user's own permissions. We test them against questions your staff write, so you know how well they work before your users rely on them.

Plan a program partnership Read about our SC26 paper

What we build

Support agents

Assistants that answer from documentation and live system data, and open a ticket for staff when a question needs a person. The ACCESS AI Assistant works this way.

MCP servers for your data

MCP (Model Context Protocol) is an open protocol for connecting AI tools to outside data. We wrap existing APIs so assistants can read allocations, system status, events or software listings.

Agent access to HPC

An Open OnDemand app that lets an assistant submit jobs, read output and manage files, running as the signed-in user under your site's existing policies.

Review assistants

Tools that check submissions against a rubric and draft feedback for a reviewer to approve. We use one for Open OnDemand Appverse submissions.

Work in production

ACCESS AI Assistant

One agent that answers ACCESS documentation questions and live questions in the same conversation, like which GPU resources are available or what is down for maintenance. It draws on 11 live data servers covering allocations, compute resources, system status, events, announcements, software, affinity groups, NSF awards, usage charts, usage data and ticketing. Supported by NSF award 2138286 under the ACCESS program.

Read the case story
Illustration: a researcher asks whether anything is down for maintenance and how to get time on Delta. The ACCESS AI Assistant answers that there are no active outages and that Delta GPU time comes through ACCESS Allocations, starting at Explore. Labels: live system status, ACCESS docs, one answer. (open full size)

OOD API for Open OnDemand

An Open OnDemand app that lets researchers and their AI assistants submit jobs, manage files and check on their cluster without a terminal. Every request runs as the signed-in user, inside the permission model Open OnDemand already has. Site operators set local policy in a plain text file that every AI client reads before acting. It is early software.

See the OOD API code (external site)
Illustration: a researcher asks an AI assistant to rerun an analysis on the GPU queue with 4 hours, and it replies that it submitted job 48213 as jsmith to the gpu partition. Labels underneath: your permissions, follows site policy, logged. (open full size)

Tested, safe and in your control

  • Tested against your questions. Graded fact by fact on questions your staff write, before launch and every release.
  • The user's own permissions. Agents can only reach what the signed-in person can.
  • Guardrails on anything risky. Only the access a task needs, sensitive files blocked, every action logged, and a warning before anything permanent.
  • People make the call. Agents explain what they will do and wait for a person to confirm.
  • Open, with your choice of model. The code is public, and agents run on open or commercial models.

Why not build it yourself?

We can build on our work for NSF ACCESS to quickly set up the backend that informs your AI assistant.

How an engagement runs

Have budget now? A one or three week discovery sprint can map the work first.

  1. Plan it into your award

    For a new proposal we provide budget text and a statement of work. Funded programs bring us on as a contractor.

  2. Build with your team

    We connect your documentation and systems, and release early to a pilot group of your users.

  3. Evaluate and report

    Test results become numbers for your annual reports.

  4. Run it for the award

    Hosting, updates and tuning, every year.

Paper

HUST workshop at SC26 · Accepted, November 2026

From Answering to Acting: An AI Agent for Research Computing User Support

A. Pasquale, V. Gazula, J. E. Bacal

How the ACCESS AI Assistant combines documentation search with live data from MCP servers, and how we test its answers.

Bring an AI assistant to your users.

Tell us what your users ask most and where your data lives. We can write the work into your next proposal, or start under your current award.

Talk to us Planning a proposal?