kodx 0.2.0

Docker-based code execution environment with LLM integration

Kodx

Codex-inspired background agent that runs on your machine

Installation

uv pip install kodx

[!IMPORTANT] Set up your API key:

export ANTHROPIC_API_KEY="your-api-key-here"

Kodx also requires Docker to be installed and running on your system.

Quick Start: kodx code

# Fix a bug or implement a feature - kodx creates a git branch automatically
kodx code "fix the failing authentication tests"

# Work from a specific branch
kodx code "add rate limiting to the API" --base-ref feature/api-v2

# Debug with uncommitted changes included
kodx code "help me fix this error" --dirty

What happens:

1. AI works in isolation - Code is written and tested in a Docker container
2. AI decides what to commit - The AI makes explicit decisions about when and what to commit
3. Git fetches committed changes - Only changes the AI commits are retrieved
4. Branch created if needed - A new branch appears in your repo only if the AI made commits

How Kodx Works

Kodx is built on a simple but powerful architecture:

The Core: kodx

At its heart, Kodx provides a flexible command (kodx) that runs AI coding assistants in Docker containers. The AI interacts with code through a shell interface, keeping your system safe while enabling powerful automation.

# The core command gives you full control
kodx --prompt "Create a web server" --export-dir ./output

See docs/kodx-core.md for advanced usage.

Specialized Commands

Built on top of the core, Kodx provides two specialized commands:

• kodx code - Git-integrated workflow that automatically creates branches and commits
• kodx ask - Read-only code analysis that never modifies files

Key Architecture

• Container Isolation: Each session runs in a fresh Docker container
• Minimal Interface: AI uses just 2 tools - shell interaction and session reset
• Git Integration: kodx code manages branches, commits, and exports automatically
• Safety First: Your working directory is never touched directly

Additional Command: kodx ask

For code analysis without modifications, use kodx ask:

kodx ask "What does this project do?"
kodx ask "Find potential security issues" --cost-limit 5.0

Like kodx code, it's a thin wrapper around the core kodx command but configured for read-only analysis.

Configuration and Customization

kodx init - Project Setup

Initialize project-specific configurations:

kodx init

This creates default .kodx/ask.yaml and .kodx/code.yaml files that customize the behavior of the specialized commands while leveraging the core kodx infrastructure.

[!TIP] Commit the .kodx/ files in your git repo and customize the behavior of kodx code.

# .kodx/code.yaml
model:
  name: claude-sonnet-4-20250514
  provider: anthropic
  max_iterations: 100

docker:
  image: node:18
  setup_script: |
    npm install -g typescript
    apt-get update && apt-get install -y git
  disable_network_after_setup: false

prompts:
  system: |
    You are a TypeScript development assistant.
    Use the provided tools to analyze and modify code.

[!TIP] Kodx's YAML configuration is built on LLMProc, allowing you to easily extend it with MCP tools.

• Learn more at https://github.com/cccntu/llmproc

Documentation

• Getting Started Guide
• kodx Core Documentation
• kodx code Documentation
• kodx ask Documentation
• Architecture & Internals
• GitHub Actions Examples

License

Apache License 2.0

Acknowledgements

Kodx is directly inspired by ChatGPT Codex. The tool design and pty-shell server design is copied from ChatGPT Codex.

Related Projects

• LLMProc - Kodx is built on top of LLMProc, a LLM process management framework.
• AIShell - AIShell is a transparent shell wrapper that provides clean context for AI.