home / mcp / oumi mcp server

Oumi MCP Server

Provides MCP access to Oumi’s YAML training configs and guided ML workflows for model fine-tuning and evaluation.

Installation

Add the following to your MCP client configuration file.

Configuration

{
  "mcpServers": {
    "aniruddh-alt-oumi-mcp": {
      "command": "oumi-mcp",
      "args": [],
      "env": {
        "OUMI_MCP_CONFIGS_DIR": "<path-to-configs>"
      }
    }
  }
}

You can run the Oumi MCP Server to give AI coding assistants access to Oumi’s library of ~500 ready-to-use YAML configs for fine-tuning LLMs. It exposes training recipes, config content, validation, and guided ML workflows to MCP-compatible clients, helping you search, retrieve, and validate training configurations efficiently.

How to use

Connect your MCP client to the server to start exploring and using training configurations. You can search for configs by model family, task, and keywords, retrieve full YAML content, verify configurations before running, and follow end-to-end ML workflows guided by the built-in resources.

How to install

Prerequisites you need on your system: Python and a working Python environment, plus your preferred MCP client.

# Install the MCP server package with MCP support
pip install oumi[mcp]

# Run the MCP server directly
oumi-mcp

# Or run as a Python module
python -m oumi_mcp_server

Configuration and usage notes

The server ships with a bundled snapshot of approximately 500 YAML config files. On startup, it checks for a fresh cached copy and syncs from GitHub if the cache is older than 24 hours. The resolution order is: OUMI_MCP_CONFIGS_DIR, then ~/.cache/oumi-mcp/configs, then the bundled configs. You can force a sync by deleting the cache and restarting the server.

# Force a refresh by clearing the cache and starting again
rm -rf ~/.cache/oumi-mcp
oumi-mcp

Configuration and project structure

You can customize where the server reads configs from by setting the OUMI_MCP_CONFIGS_DIR environment variable. The server is designed to work immediately after installation and can operate offline, with lazy background sync to keep configs up to date.

Project structure highlights important files such as the main server, configuration parsing, and the bundled configs directory.

Example workflow

After connecting your AI assistant, you can request a specific training configuration, retrieve the full YAML, then validate it before running.

Find me a LoRA config for fine-tuning Llama 3.1 8B on my custom dataset

Available tools

get_started

Overview of capabilities and quickstart guide for using the MCP server.

list_categories

Discover available model families and config types provided by the server.

search_configs

Find training configs by filters such as model, task, and keywords.

get_config

Retrieve details of a specific config and its full YAML content.

validate_config

Validate a config file before running to ensure it meets the required task type.