Skip to main content

Template-based MCP server for remote job execution

Project description

Make-Vibe

This project is designed to allow an AI coding agent to tune a code while performing performance experiments on a remote HPC system. You will be able to edit and develop code locally in a vibe coding tool while compiling and running on remotely. This includes support for doing the execution on multiple nodes.

How It Works

  • Prepare the code you would like to work on with an associated Makefile.
  • Prepare a template for a jobscript on your remote machine.
  • An MCP server is dynamically generated from your Makefile and exposes the targets to your local LLM.
  • BATCH mode: Every time a tool call is made, the code is synchronized to the remote machine, a jobscript is generated and submitted, and the results of the jobscript are returned to the LLM.
  • PERSISTENT mode: Jobs are submitted to a preallocated compute node.
  • LOCAL mode: commands executed on the same computer as the vibe coding tool.

Quick Start

# Install
pip install -e .

The example/ includes:

  • nbody - turn a naive n-body code into an openmp parallel and optimized code.
  • matmul - optimize the compiler flags for a matrix multiplication code.
  • heat1d - MPI porting of a heat flow calculation.

Key Features

  • Automatic tool generation - Parses your Makefile, creates MCP tools for each target.
  • Remote execution - Built on remotemanager for job submission
  • Resource control - Specify MPI/OMP parameters per execution
  • Build / run variables - Pass compiler flags and other env vars (e.g., cflags="-O3")

Main Components

Makefile Target Definition:

...
# @env: CFLAGS - Compiler flags      ← Becomes MCP tool parameter
build:                               ← Becomes tool name: make_build
    $(CC) $(CFLAGS) -o app app.c     ← Actual command
...

Jobscript template:

...
#SBATCH --cpus-per-task=#OMP:default=1#      ← Becomes MCP tool parameter
#SBATCH --ntasks=1
module load system/fx700 FJSVstclanga/1.0.30.01
export OMP_NUM_THREADS=${SLURM_CPUS_PER_TASK}

Template Configuration:

{
  ...
  "submitter": "sbatch",                          ← Jobscript submitter
  "template_params": {"MPI": "How many mpi..."},  ← Becomes tool parameters
  "template_file": "fx700.sh"                     ← Job script template file
  ...
}

Run Modes

The code can run in three different modes. PERSISTENT and BATCH are the modes this tool accomplishes. LOCAL could be useful if you want to enforce some fixed behavior / environment setting procedures for an LLM.

┌───────────────────────────────────────────────────────────────────────────┐
│          LOCAL: Developer's Machine (Claude Code + MCP Server)            │
│                                                                           │
│          ┌──────────────┐         ┌─────────────────┐                     │
│          │  Makefile    │────────>│   make-vibe     │                     │
│          │  (targets)   │  parsed │   MCP Server    │                     │
│          │              │         │                 │                     │
│          │ • build      │         │  Auto-generates │                     │
│          │ • run        │         │  MCP tools from │                     │
│          │ • run-large  │         │  targets        │                     │
│          └──────────────┘         └─────────────────┘                     │
│                 +                          |                              │
│          ┌──────────────┐                  |                              │
│          │  Template    │                  v                              │
│          │  config.json │         ┌─────────────────┐                     │
│          │  template.sh │────────>│  Tool Call      │                     │
│          │              │         │  + Parameters   │                     │
│          │ • host       │  fills  │  (MPI=4, etc.)  │                     │
│          │ • submitter  │  params └─────────────────┘                     │
│          │ • parameters │                  |                              │
│          └──────────────┘                  | + rsync                      │
└────────────────────────────────────|──────────────────────────────────────┘
                                     |
              ┌──────────────────────┼────────────────────────┐
              |                      |                        |
              v                      v                        v

┌──────────────────────┐  ┌──────────────────────┐  ┌──────────────────────┐
│  MODE 1: LOCAL       │  │  MODE 2: PERSISTENT  │  │  MODE 3: BATCH       │
│  submitter: "bash"   │  │  submitter: "srun    │  │  submitter: "sbatch" │
│                      │  │    --jobid=XXX bash" │  │                      │
├──────────────────────┤  ├──────────────────────┤  ├──────────────────────┤
│                      │  │                      │  │                      │
│  ┌────────────────┐  │  │  ┌────────────────┐  │  │  ┌────────────────┐  │
│  │ Local Machine  │  │  │  │ Pre-allocated  │  │  │  │  Job Queue     │  │
│  │                │  │  │  │ Compute Node   │  │  │  │                │  │
│  │  Execute       │  │  │  │                │  │  │  │ Job 1→[SLURM]  │  │
│  │  directly      │  │  │  │  [Active job]  │  │  │  │ Job 2→[SLURM]  │  │
│  │                │  │  │  │   Submit to    │  │  │  │ Job 3→[SLURM]  │  │
│  │  bash script   │  │  │  │   this job     │  │  │  │                │  │
│  │                │  │  │  │                │  │  │  │  [Each call =  │  │
│  └────────────────┘  │  │  │ [Job keeps     │  │  │  │  New job]      │  │
│                      │  │  │  running]      │  │  │  │                │  │
│                      │  │  └────────────────┘  │  │  └────────────────┘  │
├──────────────────────┤  ├──────────────────────┤  ├──────────────────────┤
│ ✓ No queue           │  │ ✓ No queue wait      │  │ ✓ Fair scheduling    │
│ ✓ Low latency        │  │ ✓ Rapid iteration    │  │ ✓ Long runs          │
│ ✓ Good for testing   │  │ ✓ Interactive feel   │  │ ✓ Independent runs   │
│                      │  │                      │  │ ✓ Cluster policies   │
│ ✗ Fixed resources    │  │ ✗ Must pre-allocate  │  │                      │
│ ✗ Must run vibe tool │  │ ✗ Fixed resource     │  │ ✗ Queue delays       │
│   on a compute node  │  │ ✗ size (#nodes)      │  │ ✗ Slower iteration   │
└──────────────────────┘  └──────────────────────┘  └──────────────────────┘

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

make_vibe-0.1.4.tar.gz (13.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

make_vibe-0.1.4-py3-none-any.whl (11.7 kB view details)

Uploaded Python 3

File details

Details for the file make_vibe-0.1.4.tar.gz.

File metadata

  • Download URL: make_vibe-0.1.4.tar.gz
  • Upload date:
  • Size: 13.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.12

File hashes

Hashes for make_vibe-0.1.4.tar.gz
Algorithm Hash digest
SHA256 1514ec59f207f87ac3f51f3ce28601e04b0d44cb4b5536304cbdac561148bd71
MD5 0599fd8a8522c69f9e7edc9766e2d0e6
BLAKE2b-256 629dda149354f9101edc21a06d7c269a032d5438cb2c656991327317a85f804c

See more details on using hashes here.

File details

Details for the file make_vibe-0.1.4-py3-none-any.whl.

File metadata

  • Download URL: make_vibe-0.1.4-py3-none-any.whl
  • Upload date:
  • Size: 11.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.12

File hashes

Hashes for make_vibe-0.1.4-py3-none-any.whl
Algorithm Hash digest
SHA256 e3f00e52971ae3bc070046a9d0964af95ff71ad6d164f92fa1a91cdbfd1e464c
MD5 4deb0368fb852813e752248ebf1ebdba
BLAKE2b-256 802df45477967c88d0d026ac3748008cb47712ee416493b14229f10e0aae6456

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page