Conflux is a tool that orchestrates autonomous development by AI coding agents based on specification-driven development. Without requiring continuous human supervision, it keeps changes moving through a full workflow: application, acceptance judgment, archiving, and final merge.

The goal is not one-off code generation. It is to define the specification first, then continuously grow a production-minded, substantial finished product by stacking changes that follow that specification.

Conflux is also not tied to any specific AI vendor. It is designed so you can swap tools such as Claude Code, Codex, and OpenCode.

• Autonomous development that keeps moving while you sleep: Even without constant human attention, AI agents process changes one by one and keep development moving forward.
• Specification-driven development: Using OpenSpec, you define the specification first, then proceed with implementation, acceptance, and improvement based on it.
• Continuously growing a substantial finished product: Instead of stopping at one-off generation, Conflux accumulates changes over time and steadily moves closer to a finished product.

• Multi-layer Ralph loops: Conflux improves through repeated iteration while keeping the context handed off in each iteration as small as possible, making LLM usage more efficient.
• Parallel development with git worktree: By assigning an independent worktree to each change, Conflux enables multiple changes to proceed safely in parallel.
• Vendor-independent agent choice: Conflux is not locked to any specific vendor such as Claude Code, Codex, or OpenCode. You can swap implementation and evaluation agents depending on the task.
• Separation of implementation and acceptance roles: By separating the role that drives implementation from the role that evaluates the result, you can combine a fast coder with a smarter reviewer. This improves overall development speed while using LLMs more efficiently.

In short, Conflux is an orchestrator for running autonomous, specification-driven development as a practical development workflow with parallel execution and clear role separation, continuously pushing a substantial finished product forward.

For server mode, remote TUI, REST API, and cflx service, see the Server Mode Guide (English).

For initial setup, see QUICKSTART.md.

# TUI
cflx

# Headless execution
cflx run

# Run only a specific change
cflx run --change add-feature-x

# Initialize the configuration file
cflx init

# Install bundled skills
cflx install-skills

The configuration file format is JSONC.

• .cflx.jsonc
• ~/.config/cflx/config.jsonc
• --config <PATH>

Generate templates:

cflx init
cflx init --template opencode
cflx init --template codex
cflx init --force

For detailed configuration examples, hooks, workspace execution, and command queue explanations, see the English README.

cargo install cflx

MIT