MLE kit MCP

MCP server providing practical tools for ML engineering workflows, including local/remote bash, a text editor, file search, remote GPU helpers (via vast.ai), and an OpenRouter LLM proxy.

Features

• bash: Run commands in an isolated Docker container mounted to your WORKSPACE_DIR.
• text_editor: View and edit files and directories in your workspace with undo support.
• glob / grep: Fast filename globbing and ripgrep-based content search.
• remote_bash / remote_text_editor / remote_download: Execute and edit on a remote GPU machine and sync files to/from it.
• llm_proxy_local / llm_proxy_remote: Launch an OpenAI-compatible proxy backed by OpenRouter locally (in the bash container) or on the remote GPU.

Requirements

• Python 3.12+
• Docker daemon available (for bash tool)
• ripgrep (rg) installed on the host (for grep tool)
• WORKSPACE_DIR should be set with a path to working directory
• Optional (for remote GPU tools): a VAST_AI_KEY with billing set up on vast.ai
• Optional (for LLM proxy tools): an OPENROUTER_API_KEY

Install

Using uv (recommended):

uv sync

Or standard pip install:

python -m venv .venv && . .venv/bin/activate
pip install -e .

Run the MCP server

Set a workspace directory and start the server. The MCP endpoint is served at /mcp.

WORKSPACE_DIR=/absolute/path/to/workdir uv run python -m mle_kit_mcp --port 5057

Defaults:

• PORT defaults to 5057 if --port is not provided
• mount_path=/ and streamable_http_path=/mcp

Claude Desktop config

{
  "mcpServers": {
    "mle_kit": {
      "command": "python3",
      "args": [
        "-m",
        "mle_kit_mcp",
        "--transport",
        "stdio"
      ]
    }
  }
}

Tools overview

• bash(command, cwd=None, timeout=60): Runs inside a python:3.12-slim container with your workspace bind-mounted at /workdir. State persists between calls. Timeouts return a helpful message.
• text_editor(command, path, ...): Supports view, write, append, insert, str_replace (with optional dry_run), and undo_edit. Only relative paths under the workspace are allowed.
• glob(pattern, path=None): Returns matching files under the workspace (optionally under path), sorted by modification time.
• grep(pattern, path=None, glob=None, output_mode=..., ...): ripgrep wrapper. Install rg on the host to enable. Output modes: files_with_matches, content, count.
• remote_bash(command, timeout=60): Runs commands on a remote vast.ai instance. Manages lifecycle unless you supply an existing instance (see env vars below).
• remote_download(file_path): Copies a file from the remote (/root/<file_path>) to your workspace.
• remote_text_editor(...): Same API as text_editor, but syncs the file(s) before and after edits to the remote.
• llm_proxy_local() / llm_proxy_remote(): Starts a small FastAPI OpenAI-compatible server backed by OpenRouter, returning a JSON string with url and scope.

Configuration (env vars)

All variables can be placed in a local .env file or exported in your shell.

• WORKSPACE_DIR (required): Absolute path to your workspace directory.
• PORT (optional): Default server port (defaults to 5057).

Remote GPU (vast.ai):

• GPU_TYPE (default: RTX_3090)
• DISK_SPACE (GB, default: 300)
• EXISTING_INSTANCE_ID (optional): Use an existing vast.ai instance instead of creating a new one.
• EXISTING_SSH_KEY (optional): Path to an SSH private key to use with the existing instance.
• VAST_AI_KEY (optional but required to launch new instances)

OpenRouter proxy:

• OPENROUTER_API_KEY (optional but required for proxy tools)
• OPENROUTER_BASE_URL (default: https://openrouter.ai/api/v1)

Notes:

• The remote GPU helper will generate an SSH key at ~/.ssh/id_rsa if one is missing, and attach it to the instance.
• Creating/destroying instances may incur cost; be mindful of environment defaults.

Development

Run tests:

make test

Lint / type-check / format:

make validate

Docker

You can also build and run via the provided Dockerfile:

docker build -t mle_kit_mcp .
docker run --rm -p 5057:5057 \
  -e PORT=5057 \
  -e WORKSPACE_DIR=/workspace \
  -v "$PWD/workdir:/workspace" \
  mle_kit_mcp