Skip to main content
Use kumo-tuning-agent when you want Codex to handle the full Kumo SDK workflow for you. It is meant for tasks where the agent should inspect the data, build the graph, validate the predictive query, train a model, tune it, and then run predictions.

Download Skill

Kumo SDK Docs

This is the better fit when the work sounds like “build and improve a model,” not just “score something quickly.” For example, this skill is a good choice if you want Codex to connect to Snowflake, inspect customer and order tables, build a churn model, and improve it until the evaluation metric looks reasonable. It also fits jobs like forecasting 30-day revenue, comparing model plans, or writing batch predictions to S3 after training is complete.

​
Install

Download https://raw.githubusercontent.com/kumo-ai/kumo-skills-catalog/master/kumo/kumo-tuning-agent/SKILL.md, then install it into Codex: 
mkdir -p ~/.codex/skills/kumo-tuning-agent
curl -L https://raw.githubusercontent.com/kumo-ai/kumo-skills-catalog/master/kumo/kumo-tuning-agent/SKILL.md -o ~/.codex/skills/kumo-tuning-agent/SKILL.md

​
Before You Start

Set these environment variables before starting Codex: 
export KUMO_API_ENDPOINT="https://your-deployment.kumoai.cloud/api"
export KUMO_API_KEY="your-kumo-api-key"

​
Example Prompts

Use the kumo-tuning-agent skill with the dataset at s3://acme-ml-datasets/churn-demo/ to inspect the schema, build the graph, and train a churn model.

Use the kumo-tuning-agent skill on the retail dataset at s3://acme-ml-datasets/retail-forecasting/ to predict 30-day revenue per customer and improve the model plan if RMSE is weak.

Use the kumo-tuning-agent skill on the fraud dataset at s3://acme-ml-datasets/credit-card-fraud/ to train a model and tell me what to improve if recall is weak.
The skill is opinionated in a useful way: it pushes Codex to inspect the real schema first, validate each step, and come back with concrete outputs such as metrics, prediction locations, and sample rows instead of a vague summary. If you want more detail on model-plan settings or evaluation, these docs are the next place to go: