mock-gkn-hpc

A public MCP mock of GKN's internal HPC server for Ansys / Finite Element analysis. It lets external clients exercise the full submit → poll → fetch-artifacts loop against a fake HPC — no real solver, queue, or cluster behind it.

This is a sample MCP server developed alongside the DUCTILE agentic orchestration paper. See the DUCTILE repository or the paper (DOI: TBD) for context.

Install and run

uvx mock-gkn-hpc

Or wire it into your MCP client config:

{
  "mcpServers": {
    "mock-gkn-hpc": {
      "command": "uvx",
      "args": ["mock-gkn-hpc"]
    }
  }
}

Job artifacts default to a system tempdir (gettempdir()/mock-gkn-hpc). Override with --runs-dir /custom/path or the MOCK_GKN_HPC_RUNS_DIR env var.

Tool surface

submit_ansys_run mirrors GKN's qansys wrapper:

All parameters round-trip through meta.json, the log banner, and the rendered results.rst.

Other tools and resources:

The log is built from the job: the requested version and product, every file the deck reads, and one LOAD STEP line per SOLVE the deck executes, with its load file.

Jobs are stateless: a job's status, log and results follow from its submit time (pending for ~30 s, then ready, expired 1 h later), so they need no running process and survive a server restart (e.g. when a client session ends and a new one starts).

Input pre-flight check (v0.6.1)

A real solve stops at the first /INPUT or CDREAD whose file it cannot read. The mock has no solver, so submit_ansys_run walks the runscript at submission (parameters, arrays, *DO loops, macros and %...% substitution) and checks every file the deck would read. A job sees:

• its staged directory, the input file's folder (relative paths resolve against it; an absolute path to the submitting machine is not visible to the job, even inside that folder), and
• the cluster's shared /project storage, described by src/mock_gkn_hpc/cluster_files.txt (override with the MOCK_GKN_HPC_CLUSTER_FILES env var, one absolute path per line).

If any file is missing, the job is recorded as failed (its log holds the error) and the tool returns status: "failed" with Ansys-style messages (a normal result, not an MCP error), e.g.

Job e816a8ac (runscript) failed during input processing: 6 file(s) read by the deck do not exist.
 *** ERROR ***
 /INPUT failed (runscript line 139). File /project/.../01_inputs/loads/limit_load_2.inp does not exist.

so an agent sees the failure in its context and can fix the deck and resubmit. A successful submission reports input_files_checked (and notes, e.g. a character value cut to 32 characters).

The walker follows APDL's parameter rules:

• array parameters must be dimensioned with *DIM before values are assigned, unless an implied (colon) loop defines them (a(1:3)=1,2,3);
• CHAR array elements hold 8 characters and scalar character parameters 32;
• in a STRING array the first subscript is the character position and the next ones select the string: name(1,j) is the j-th string. Writing name(2)='x' into a one-column STRING array therefore writes from character 2 of the same string.

It also handles *SET, NINT, INT, ABS, CHRVAL and STRCAT. A /INPUT or CDREAD whose file name cannot be resolved with this subset is reported as a deck error, not skipped.

Each client can keep its own job store with --runs-dir (useful when several agents run in parallel and should not see each other's jobs in list_jobs).