miliario 0.2.0

miliario — self-hosted ML experiment tracking & LLM observability (imports as `miliario`; the historical `mlop` import keeps working)

miliario Python client

The Python client for miliario — a self-hosted ML experiment tracker & LLM/agent observability platform:

pip install miliario

import miliario

The historical mlop-sl distribution is an alias of miliario, and the historical import mlop / mlop CLI / MLOP_* environment variables keep working as thin compatibility shims — existing code needs no changes.

Full usage docs: https://slenser0.tech/docs (raw markdown for AI agents: /docs.md).

Install from source

git clone https://github.com/SLENSER0/ml_logger
cd ml_logger
pip install .

MCP server (Claude Code / Cursor)

miliario mcp runs a read-only MCP stdio server over your miliario instance, so coding agents can answer questions like "which run had the best val_acc?" or "why did eval costs spike yesterday?".

pip install 'miliario[mcp]'

Register with Claude Code:

claude mcp add miliario -e MILIARIO_API_KEY=mlps_... -e MILIARIO_HOST=https://slenser0.tech -- miliario mcp

Or in a generic MCP client config (Cursor mcp.json and similar):

{
  "mcpServers": {
    "miliario": {
      "command": "miliario",
      "args": ["mcp"],
      "env": {
        "MILIARIO_API_KEY": "mlps_...",
        "MILIARIO_HOST": "https://slenser0.tech"
      }
    }
  }
}

If MILIARIO_API_KEY is unset, the server falls back to the token stored by miliario login. Tools: list_projects, list_runs, list_metrics, get_metric_series, search_spans, get_costs, get_run_url.

Original repo: mlop-ai/mlop