Akkadian

MCP server for ML competition workflows. Provides AI-agent-friendly tooling for Kaggle competitions with integrated experiment tracking (MLflow) and hyperparameter tuning (Optuna).

Install

pip install akkadian                    # Core (Kaggle + knowledge graph)
pip install akkadian[mlflow]            # + MLflow tracking
pip install akkadian[optuna]            # + Optuna tuning
pip install akkadian[all]               # Everything

Quick Start

Add to your project's .mcp.json:

{
  "mcpServers": {
    "akkadian": {
      "type": "stdio",
      "command": "akkadian-mcp",
      "args": ["--working-dir", "."]
    }
  }
}

Then use the bootstrap init tool to configure your project.

Tools

Tool	Operations	Requires
bootstrap	init, status, update_config	core
kaggle	status, list_kernels, submissions, push_kernel, download_output	core
knowledge	status, search, entity, neighbourhood, list, record_score, add_entity, add_relation	core
mlflow	setup, experiments, runs, compare, best, ingest_artifact, suggest	akkadian[mlflow]
optuna	create_study, studies, trials, best, importance, suggest_space, prune_config	akkadian[optuna]

How It Works

Akkadian stores all state in a .akkadian/ directory in your project root:

• config.json — project configuration (competition, username, goal metric)
• knowledge.db — SQLite knowledge graph tracking models, kernels, datasets, scores
• mlflow.db — MLflow experiment tracking (if installed)
• optuna.db — Optuna study storage (if installed)

The agent adapts to your workflow — it observes, tracks artifacts, and suggests strategies without enforcing a specific process.

Development

git clone https://github.com/manwithacat/akkadian.git
cd akkadian
pip install -e ".[dev,all]"
pytest

License

MIT