kumo-coding-agent

Kumo Coding Agent — context docs, skills, and self-maintenance for building ML models with the Kumo SDK

public

Forked

Kumo Coding Agent

A portable collection of context and skills that extends any LLM coding tool with deep knowledge of the Kumo ML platform. Works with Claude Code, Codex, Cursor, and any tool that reads markdown.

Setup

Prerequisites

kumoai: pip install kumoai (needed for running predictions, not for the agent itself)
GitHub CLI: brew install gh && gh auth login (needed only for /kumo-issue and /kumo-pr commands)

Claude Code

1. Add the agent to your project:

cd your-project
git clone https://github.com/kumo-ai/kumo-coding-agent.git

2. Tell Claude Code to read the agent:

echo 'Also read kumo-coding-agent/CLAUDE.md for Kumo agent capabilities.' >> CLAUDE.md

3. Set up permissions (zero-prompt setup):

bash kumo-coding-agent/setup.sh

4. Install slash commands (optional):

npx skills add kumo-ai/kumo-coding-agent --agent claude-code

This installs /kumo-issue and /kumo-pr for reporting bugs and contributing fixes.

5. Start using it. Ask questions in natural language:

"Predict which customers will churn in the next 30 days by running RFM on the SALT dataset"

Codex

1. Add the agent to your project:

cd your-project
git clone https://github.com/kumo-ai/kumo-coding-agent.git

2. Tell Codex to read the agent:

echo 'Also read kumo-coding-agent/CLAUDE.md for Kumo agent capabilities.' >> AGENTS.md

3. Set up permissions (zero-prompt setup):

bash kumo-coding-agent/setup.sh

4. Install slash commands (optional). Inside a Codex session:

$skill-installer install https://github.com/kumo-ai/kumo-coding-agent

5. Start using it. Ask questions in natural language.

Cursor

1. Add the agent to your project:

cd your-project
git clone https://github.com/kumo-ai/kumo-coding-agent.git

2. No extra setup needed. Cursor reads .cursor/rules/ automatically.

3. Install slash commands (optional):

npx skills add kumo-ai/kumo-coding-agent --agent cursor

4. Start using it. Ask questions in natural language.

Other tools

Clone the repo into your project. Any LLM tool that reads markdown can use CLAUDE.md as the entry point.

git clone https://github.com/kumo-ai/kumo-coding-agent.git

Updating

To pull the latest version of the agent:

cd kumo-coding-agent && git pull

To pin to a specific release:

cd kumo-coding-agent && git fetch --tags && git checkout v1.0.0

What Can It Do?

Task	Start Here
Build a prediction model end-to-end	`skills/scope-prediction-task.md`
Get instant predictions (no training)	`skills/rfm-predict.md`
Write or fix a PQL query	`skills/write-pql.md`
Train a high-accuracy model	`skills/train-model.md`
Debug a failed prediction	`skills/debug-prediction.md`
Improve weak model performance	`skills/iterate-model.md`
Design a SQL+PQL business workflow	`context/patterns/prediction-patterns.md`
Decide between RFM and training	`context/guides/rfm-vs-training.md`

Commands

Two commands are available for reporting issues and contributing fixes.
Requires GitHub CLI (gh auth status must pass).

Command	What it does
`/kumo-issue [description]`	Files a GitHub issue on this repo
`/kumo-pr [description]`	Creates a branch, fixes a doc, and opens a PR

Directory Structure

kumo-coding-agent/
├── CLAUDE.md              # Entry point for Claude Code
├── AGENTS.md              # Entry point for Codex (points to CLAUDE.md)
├── .cursor/rules/         # Entry point for Cursor (points to CLAUDE.md)
├── context/               # Curated Kumo knowledge (loaded on demand)
│   ├── platform/          # SDK, RFM, PQL, graph, connectors
│   ├── guides/            # Decision guides
│   ├── patterns/          # Business workflow patterns
│   └── verticals/         # Industry-specific guides
├── skills/                # Step-by-step workflows
├── meta/                  # Add docs, sync from upstream, verify
├── scratch/               # Session state (gitignored)
└── eval/                  # Test questions for agent quality

Extending the Agent

Action	Instructions
Add a context document	`meta/skills/add-context-doc.md`
Add a workflow skill	`meta/skills/add-skill.md`
Sync from upstream repos	`meta/skills/sync-from-source.md`
Check doc freshness	`meta/skills/validate-freshness.md`
Verify claims against code	`meta/skills/verify-content.md`
Audit known gaps	`meta/skills/check-gaps.md`

Design Principles

Portable: Plain markdown, works with any LLM tool
On-demand loading: Only loads what the task needs
Version-tracked: Every doc traces to a specific package version
Self-correcting: Gap manifest + verify-content catch drift
Testable: Eval questions verify knowledge quality

We love your feedback! ❤️

As you work with the Kumo Coding Agent, if you encounter any problems or
things that are confusing or don’t work quite right, please
open a new issue.
Join our Discord!

Community contribution 🤝

If you’re considering contributing an example notebook or a new skill,
please first open a new issue
and describe your proposal so we can discuss it together before you invest a
ton of time. We’ll invite you to our Mountain View, CA office (if you’re local)
or send you Kumo Swag if your contribution is accepted.

Thank you and excited to see what you’ll build with the Kumo Coding Agent!

License

Released under the MIT License.

Find me

v0.3.3[beta]