jobstate

Tools for AI agents to submit, watch and report on Slurm jobs on the JSC supercomputers JUPITER, JURECA, JUWELS and JUSUF: an MCP server (jobstate-mcp) and a command-line tool (jobstate).

Instead of polling squeue turn after turn, an agent waits for its jobs with one tool call, or is notified when they end; a long job costs one model turn at the end. Submissions go through a dry run and a separate confirmation, under limits and budget guards the server enforces itself.

Install

pip install jobstate

Or with uv: uv tool install jobstate.

On a JSC login node, check the setup and add the server to Claude Code:

jobstate check
claude mcp add -s user jobstate -- jobstate-mcp

From your own machine the server reaches the cluster over ssh:

jobstate check --cluster jupiter --host <your ssh config name>
claude mcp add -s user jobstate-jupiter -- jobstate-mcp --cluster jupiter --host <your ssh config name>

Tools

job_status, list_jobs, job_summary, tail_log, job_report (GPU efficiency from LLview), watch_job, wait_for_jobs, cluster_info, and, unless configured read-only, submit_job (a dry run), confirm_submit and cancel_job. End-of-job events carry exit code, run time against the limit, busy cores, peak memory, I/O and node-hours from Slurm accounting.

Configuration (allowed accounts, tighter limits, read-only mode, ntfy or webhook notifications) goes into ~/.config/jobstate-mcp/config.yaml. Python 3.12 or later.