Codex-inspired background agent that runs on your machine
uv pip install kodxImportant
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.
# 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" --dirtyWhat happens:
- AI works in isolation - Code is written and tested in a Docker container
- AI decides what to commit - The AI makes explicit decisions about when and what to commit
- Git fetches committed changes - Only changes the AI commits are retrieved
- Branch created if needed - A new branch appears in your repo only if the AI made commits
Kodx is built on a simple but powerful architecture:
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 ./outputSee docs/kodx-core.md for advanced usage.
Built on top of the core, Kodx provides two specialized commands:
kodx code- Git-integrated workflow that automatically creates branches and commitskodx ask- Read-only code analysis that never modifies files
- 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 codemanages branches, commits, and exports automatically - Safety First: Your working directory is never touched directly
For code analysis without modifications, use kodx ask:
kodx ask "What does this project do?"
kodx ask "Find potential security issues" --cost-limit 5.0Like kodx code, it's a thin wrapper around the core kodx command but configured for read-only analysis.
Initialize project-specific configurations:
kodx initThis 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
- Getting Started Guide
kodxCore Documentationkodx codeDocumentationkodx askDocumentation- Architecture & Internals
- GitHub Actions Examples
Apache License 2.0
Kodx is directly inspired by ChatGPT Codex. The tool design and pty-shell server design is copied from ChatGPT Codex.