hpc-bridge

A batch supercomputer as an AI agent's REPL

Gus Ellerm · Ryan Chard · September 2026

gusellerm@uchicago.edu

QR code linking to github.com/ryanchard/hpc-bridgegithub.com/ryanchard/hpc-bridge

What hpc-bridge is

A Claude Code plugin that gives an agent a REPL on HPC.

  • An MCP server plus a skill. Four verbs as tools — discover, connect, run, release — and the facility knowledge an agent cannot infer on its own.
  • It steers the agent onto a happy path. Every step is one tool call with a typed outcome, so the easy route is the right one.
  • It fences off the rest. (Ideally) no ssh into a login node, no job scripts or queue watching, no guessing each facility's quirks.
  • It asks before it spends, and reports honestly. Nothing is billed without a question; draining is never reported as down.

Same theme, yesterday: Haotian Xie, Diamond Agent · Tanjin He, User Experience Sharing · Parmanand Sinha, No SSH Required · Geoffrey Lentner, The Anvil MEP

hpc-bridge·Gus Ellerm · Ryan Chard · September 2026

Why do we want agents on supercomputers?

facility augment a researcher you your laptop agent minutes per turn run a research campaign you goal · budget hosted · cloud or facility agent hours to days still unsolved for a hosted agent identity for a non-human actor service accounts, delegated consent budget caps and revocation audit: who asked for what
hpc-bridge·Gus Ellerm · Ryan Chard · September 2026

Discovery and access

you log in once organisation Globus Search index your org decides who sees what facility runs as your account discover run · zero SSH multi-user endpoint agent one Globus identity also discovers through cached records of earlier connects your own endpoints already online a login-node probe, for an un-indexed cluster
hpc-bridge·Gus Ellerm · Ryan Chard · September 2026
endpoint reconnect teardown · detach attach reuse create first task unbound online offline live block · your Slurm/PBS job canary stop idle next call cancelled rejected · no account · conflict provisioning up draining cold_start down 1 2 3 4 5
hpc-bridge·Gus Ellerm · Ryan Chard · September 2026

Measured: a REPL turn on hpc-bridge

Where a REPL turn's time goes (mean)agent time — choosing the next steptool time — running it0 s1 s2 s3 s4 s5 slocal Bash tool1.71 s1.70 shpc-bridge · fake cluster4.71 s2.65 shpc-bridge · Globus Cluster4.78 s2.69 sTool latency per step (every warm step, every run)0 s1 s2 s3 s4 s5 slocal Bash toolmedian 0.01 shpc-bridge · fake clustermedian 2.03 shpc-bridge · Globus Clustermedian 1.95 s

  • A turn takes about 3 s longer than on a laptop. 1.7 s locally, 4.7 s through hpc-bridge.
  • A third of that is the agent reading more. A typed result with state and notices takes about 1 s longer to think about than bare stdout.
  • Two thirds is the trip through Globus Compute. About 2 s to submit, dispatch, run and return. The cluster itself adds nothing measurable: the simulated one is just as slow.
hpc-bridge·Gus Ellerm · Ryan Chard · September 2026