2026-11-19 17:04:57 +01:00
2026-11-19 13:56:13 +01:00
2026-11-19 13:56:13 +01:00
2026-11-19 13:56:13 +01:00
2026-11-19 13:56:13 +01:00
2026-11-19 13:56:13 +01:00
2026-11-19 14:43:16 +01:00
2026-11-19 15:25:47 +01:00
2026-11-19 13:56:13 +01:00

merlin-hpc-mcp

AI-mediated access to our Slurm clusters, with containment: curated, narrow tools; per-user identity; provenance-tagged jobs; audited end to end.

Built on Purdue RCAC's cluster-mcp as a dependency, not a fork (see ADR-0001); the upstream revision is pinned in the root pyproject.toml and governed by docs/upstream-policy.md.

Layout

merlin-hpc-mcp/
├── pyproject.toml            # uv workspace root + THE upstream pin
├── Makefile                  # sync / test / lint / check-upstream
├── .github/workflows/
│   ├── ci.yml                # lint, format, types (advisory), test
│   └── upstream-watch.yml    # weekly advisory build against upstream main
├── packages/
│   └── merlin-hpc-mcp/    # the site package (workspace member)
│       ├── src/merlin_hpc_mcp/
│       │   ├── server.py         # assembly entrypoint (stub until M1)
│       │   ├── compat.py          # the ONE import wall vs upstream
│       │   ├── config.py          # MERLIN_HPC_MCP_* settings
│       │   ├── policy/            # allowlist + provenance (working now)
│       │   └── tools/             # site tools (from M1)
│       └── tests/
├── docs/
│   ├── architecture.md           # target shape + milestones M0-M4
│   ├── upstream-policy.md        # pinning, drift triage, upgrade checklist
│   └── decisions/                # ADRs
├── deploy/
│   ├── systemd/                  # gateway unit template (M2)
│   ├── nginx/                    # TLS front template (M2)
│   └── agents.d/                 # cluster context served to agents
└── configs/mcp-clients/          # user-facing registration examples

Status

Milestone M0 (organisation, see docs/architecture.md). Policy primitives are implemented and tested; server wiring (M1) starts with the verification spike described in compat.py.

Developing

make sync
make test
make check-upstream    # advisory: run our tests against upstream main

Principles (short version)

  1. Slurm-side fences (QOS, associations, job_submit plugin) are the only controls we fully trust; gateway policy shapes what is easy and auditable; client-side behaviour is UX, never a boundary.
  2. Every job the gateway submits is tagged --comment=ai-mcp; the cross-audit (gateway logs vs sacct) flags anything that went around.
  3. No generic shell tool for agents, ever.

License

MIT

S
Description
MCP endpoint for using Merlin compute resources with AI Agentic tools
Readme MIT
114 KiB
0 Stars 0 Watchers 0 Forks
Languages
Python 97.1%
Makefile 2.9%