Drive a supercomputer from Claude Code. Ask for a compute node in plain language; the agent finds the facility, logs you in to Globus once, starts a one-node scheduler job (a block, and it asks before it spends anything), runs your commands on that node, and releases it when you are done.
Two commands inside Claude Code; the repository is its own plugin marketplace.
/plugin marketplace add ryanchard/hpc-bridge
/plugin install hpc-bridge@hpc-bridge
Then ask "What HPC facilities can I use?" — that first answer needs no login and no configuration.
Another MCP host? hpc-bridge is a standard MCP server — point hermes-agent, Claude Desktop, Cursor, or the OpenAI Agents SDK at uvx --from git+https://github.com/ryanchard/hpc-bridge hpc-bridge. See other MCP hosts.
| You bring | Included |
|---|---|
| Claude Code | the plugin: the tools, and a skill the agent follows on its own, so plain requests work |
uv, the only local prerequisite (it fetches Python 3.11 or newer itself if you don't have one) |
the public facility registry: which machines exist and what each needs from you |
| A Globus account, logged in with the identity your facility knows you by (usually institutional or ACCESS) | the Globus login in your terminal: the browser opens once, the token is remembered |
Access to a facility: an account whose Globus identity it has mapped (zero SSH), or an account plus key-based SSH to its login node in ~/.ssh/config |
the endpoint bootstrap: on an SSH facility hpc-bridge installs Globus Compute in your home directory there, then reconnects with no SSH — nothing for you to install on the machine |
| For an un-catalogued cluster, a minute to confirm the settings the agent discovers | discovery of that cluster's configuration, the spend gate before any billed block, and idle self-release |
Nothing hpc-bridge-specific is configured by hand. Details: Install · Facilities.
- What HPC facilities can I use?
- Connect me to Anvil — a facility from step 1, or the cluster at
login.example.edufor one that isn't listed - Run
hostnameon a compute node - Release the compute block
The connect opens your browser for Globus; approve once. Before the first command the agent shows the partition and account it will charge and asks you to confirm. Allocating a block takes a couple of minutes on a quiet cluster; after that, commands run immediately and your working directory persists between them. The Quickstart walks through each step and what to expect.
SSH is a one-time bootstrap; everything after rides a Globus Compute endpoint. The connect step resolves a machine from the registry, from a cluster you used before, or by probing a new login node, and both discovery and compute then flow through the endpoint over Globus. A facility that runs its own multi-user endpoint needs no SSH at all. Design, tool reference and module notes: the vault.
- User guide — install, quickstart, facilities, the Globus login, costs and stopping, troubleshooting.
- Adding a facility to the registry — for facility staff and contributors.
- Vault — how it is built.
- Bugs and questions: GitHub issues.
Public beta (0.1.6, Apache-2.0): tested live on six clusters, security-reviewed, and now in the hands
of test users — expect rough edges and please report them. In the
registry: Purdue Anvil (multi-user endpoint), NCSA Delta (multi-user endpoint), SDSC Expanse
(SSH with a one-time code) and Globus Labs' cluster; also proven live through the bring-your-own path on Midway
(Slurm) and ALCF Polaris (PBS). Slurm and PBS are the supported schedulers.
Unit tests, lint and type checks run in CI. Every command after the bootstrap carries only a scoped Globus token,
never SSH material; SSH is key-only and used once to bootstrap. A facility that asks for a one-time code
(TOTP or Duo passcode) at login gets it from you in the chat, single-use; a password is never handled by the agent:
it hands you an ssh command for your own terminal and shares that session. The agent runs as you on the login
node, so treat what it runs there as you would your own shell.