ML Experiment Tracker MCP Server

An MCP (Model Context Protocol) server that lets assistants create experiment runs, log metrics, record hyperparameters, tag runs, and compare results. Persistence is SQLite, so your history survives across sessions.

Stack

• Python 3.10+
• Official MCP Python SDK (mcp) with FastMCP
• SQLite for experiments, tags, hyperparameters, and time-series metrics

Quick start

cd "/path/to/ml-experiment-tracker-mcp"
python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .

Run the server (stdio — used by Cursor and similar clients):

python -m ml_experiment_tracker_mcp.server

Or:

ml-experiment-tracker-mcp

Dev / inspector (requires MCP CLI extras):

pip install "mcp[cli]"
mcp dev ml_experiment_tracker_mcp/server.py

Cursor setup

In Cursor Settings → MCP, add a server:

{
  "mcpServers": {
    "ml-experiment-tracker": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["-m", "ml_experiment_tracker_mcp.server"],
      "cwd": "/absolute/path/to/ml-experiment-tracker-mcp"
    }
  }
}

Optional: set ML_EXPERIMENT_TRACKER_DB to a full path for the SQLite file (default: experiments.db in the process working directory).

Web dashboard (local)

Browse runs, metrics summaries, and side-by-side comparisons in the browser:

pip install -e ".[web]"
ml-experiment-tracker-ui
# http://127.0.0.1:8845

Override host/port with ML_EXPERIMENT_TRACKER_UI_HOST and ML_EXPERIMENT_TRACKER_UI_PORT.

Live demo (portfolio)

Public read-only dashboard with seeded sample runs:

ml-experiment-tracker-demo.onrender.com

Deploy via Render Blueprint (render.yaml + Dockerfile). Demo env:

• ML_EXPERIMENT_TRACKER_DEMO=1 — seeds vision/NLP sample runs on first boot
• ML_EXPERIMENT_TRACKER_DB=/data/experiments.db — SQLite on container disk (resets on free-tier redeploy)

Health check: GET /health

Tools

Tool	Purpose
create_experiment	New run with optional description, tags, hyperparameters
log_metric	Log a scalar (e.g. loss / accuracy), optional step
set_hyperparameter	Update one hyperparameter
add_experiment_tags	Add tags
finish_experiment	Mark completed, failed, or aborted
list_experiments	Filter by status/tag
get_experiment_detail	Full run with all metrics
compare_experiments	Side-by-side metric aggregates
delete_experiment	Remove a run

Resource: experiment://{experiment_id} — same payload as get_experiment_detail.

Resume angle

You can describe this project as: Designed and shipped an MCP server for ML experiment tracking with SQLite persistence, exposing typed tools/resources for LLM clients (Cursor, Claude, etc.).

License

MIT