mle_kit_mcp

MLE kit MCP provides a multi-tool development environment for ML workflows inside a containerized server. The server exposes utilities such as a local bash environment inside a Docker container bound to your workspace, a text editor for editing files with undo support, fast file search, remote GPU helpers via vast.ai, and an OpenRouter-backed OpenAI-compatible proxy. You can interact with these tools through the MCP endpoint served by the mle_kit_mcp module. Start the MCP server and connect to the /mcp endpoint to issue commands and leverage the provided interfaces. The tools are designed to be used both locally (within the container) and remotely (on a remote GPU) depending on the operation you choose, such as remote_bash for executing commands on a distant machine or remote_text_editor to edit files remotely and sync them back to your workspace. The llm_proxy_local and llm_proxy_remote utilities let you run an OpenRouter-backed LLM proxy either locally inside the bash container or on the remote GPU.

Once running, you can access the following capabilities:

• Bash: Run commands inside an isolated Docker container with your WORKSPACE_DIR mounted, enabling repeatable experiments and isolated environments.
• Text editor: View and edit files in your workspace with undo support, directly through the MCP interface.
• Glob / grep: Quickly glob files or search file contents using ripgrep-backed utilities; useful for code exploration and refactoring.
• Remote GPU tools: Launch remote computations on vast.ai GPUs, and synchronize files to/from the remote machine for faster data processing and model training.
• LLM proxy: Start an OpenAI-compatible proxy backed by OpenRouter, enabling local or remote LLM access with customizable routing.

Prerequisites

• Python 3.12+ installed on your host
• Docker daemon running if you plan to use the local bash tool
• ripgrep (rg) installed on the host for grep tool usage
• Network access to install Python packages (online environment)

Installation steps (recommended via uvx)

1. Create and activate a virtual environment (optional but recommended):

python -m venv .venv
source .venv/bin/activate

2. Install the MCP package in editable mode (this repo):

python -m pip install -e .

3. Install uvx (if not already installed) and ensure uv is available as a command:

pip install uvx
# or ensure uv is installed and accessible as `uvx`

4. Run the MCP server (example port 5057):

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

5. Optional Docker run (same behavior via Docker):

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

Notes:

• The server defaults to port 5057 if you do not specify --port.
• Ensure WORKSPACE_DIR is set to an absolute path where you want the workspace mounted inside the container.
• For remote GPU features, configure the required environment variables such as VAST_AI_KEY and related settings as documented in the README.

Tips and common issues:

• If you encounter permission or filesystem issues, ensure the host user has access to the workspace directory and Docker is allowed to bind mount the workspace.
• Remote GPU tools rely on vast.ai; ensure you have a valid VAST_AI_KEY and that you understand the cost implications of creating remote instances.
• For LLM proxy tools, you may need an OPENROUTER_API_KEY; set OPENROUTER_BASE_URL if you use a custom OpenRouter endpoint.
• The bash tool runs inside a Docker container with the workspace mounted at /workdir; commands should reference paths relative to the workspace, not the host filesystem.
• If you edit the server or dependencies, re-install in editable mode to pick up changes, or rebuild the Docker image.
• The MCP endpoint is served at /mcp by default; you can specify different transports if supported by your frontend client.