github.com/Zhanyl-tech/slurm-mcp · Python · MIT

built The guard, the topic surface and the stdio server are built and tested against fake Slurm binaries and hand-written fixtures, in-process and over real stdio. It has not been run against a real Slurm cluster, and the fixtures are illustrative, not recordings.

A system prompt that says “only use read-only commands” is a request, not a control. It fails open: a jailbreak, a confused tool call, or an ordinary hallucination is enough to reach scontrol update on a production controller.

The guard is an allowlist of shapes

Version 0.1.0 refused a list of forbidden subcommands, and an audit found 23 state-changing command lines that passed it: Slurm accepts unique abbreviations (scontrol upd is update), an option with a value can hide the verb, some verbs were missing from the list, and a bare scontrol reads commands from stdin. Version 0.2.0 inverts it. A command must pass two layers:

  • binary shape — one of seven read binaries, each with only the exact option spellings listed for it; scontrol only as scontrol show <config|job|node|partition> [name]; sdiag with no options, because sdiag --reset zeroes the counters a diagnosis reads;
  • topic shape — exactly one topic’s command plus that topic’s filters, whose values cannot start with - or contain whitespace, shell metacharacters or control characters.

Commands run without a shell anyway. The tests drive the guard with 70 adversarial command lines — the 23 bypasses, each with its man-page and source citation, plus mutations, off-allowlist commands, injections and off-shape arguments — and every one is refused (counted from the test corpus; the guard’s 104 tests passed on 2026-09-26, and a test keeps the README’s figure equal to the corpus). The stronger control is Slurm’s own authorization: run the server as an unprivileged Slurm user, and a guard bug that let a write through would still be refused by slurmctld.

A read-only server can still hurt a struggling controller, and squeue’s man page warns against RPC loops, so live reads are cached briefly, capped in concurrency and rate, and truncated to a row limit with a note saying so. The defaults are judgement calls, not tuned against a real controller.

Three tools, not one per binary

slurm_overview answers the question most sessions open with, slurm_query takes a topic from a closed vocabulary, and slurm_describe fetches column meanings for one topic only when needed. make footprint, run on 2026-09-26:

  resident: three tool descriptors    1486 chars
  resident: server instructions        369 chars
  resident total                      1855 chars
  detail, fetched on request          4229 chars
  flat, 7 schemas only                2009 chars  (1.08x resident)
  flat, schemas + inlined detail      6238 chars  (3.36x resident)

Read it carefully: a flat server with one schema per binary and no guidance is about the same size as this one’s resident surface. The 3.36× holds only if a flat server would carry comparable guidance in its descriptions, which is an assumption, not a measurement. What the measurement does show is that the 4,229 characters of detail stay off the resident path until asked for. (This page used to quote v0.1.0’s 1,088 against 4,441 characters; that ratio rested mostly on the unstated assumption.)

That is a context-cost measurement, not a quality claim. Whether it changes an agent’s diagnosis is unmeasured.

Not the only Slurm MCP server

Other public Slurm MCP servers exist, and the ones checked expose write paths (submit, cancel, hold, requeue). This one differs in scope: read-only by construction, a closed topic vocabulary instead of a command line, and detail disclosed on request. A slurmrestd backend, which would suit a server running next to a Slinky-managed cluster, does not exist yet.