A lightweight, model-agnostic AI coding runtime. Inspired by recent Claude Code architecture discussions.
Brane Code is a simple CLI-based AI coding runtime that:
- works with OpenRouter and OpenAI models
- runs locally
- executes multi-step coding tasks
- provides a clean alternative to existing agent-based coding tools
Recently, Claude Code’s architecture became widely visible through a source map exposure.
Most responses have been:
- direct forks
- minimal changes
- same workflows
Brane Code takes a different approach.
Instead of copying, it rebuilds the core ideas into a cleaner, model-agnostic system.
- Accepts natural language instructions
- Plans and executes coding tasks
- Reads, writes, and edits files
- Runs commands and tools
- Works across multiple LLM providers
All inside a simple terminal interface.
No lock-in.
- OpenAI
- OpenRouter
- (extensible to others)
Designed to be:
- fast
- local-first
- minimal overhead
Instead of single prompts, Brane Code:
- breaks problems into steps
- uses tools
- iterates toward a solution
brane fix "auth token bug"Brane Code will:
- inspect relevant files
- plan the fix
- apply changes
- return a diff + explanation
Early version.
Right now, Brane Code is a:
clean, working alternative to Claude Code-style runtimes
This is just the starting point.
The direction is clear:
- persistent memory across sessions
- replayable AI coding workflows
- structured, inspectable agent runs
- systems that improve over time
Most AI coding tools today are:
- stateless
- non-reproducible
- hard to inspect
Brane Code is moving toward:
AI development that is persistent, structured, and compounding
Brane Code is an independent project inspired by publicly discussed architectures. It is not affiliated with or endorsed by Anthropic.
Brane Code starts as a simple alternative.
It is being built toward something much larger:
a system where AI coding work doesn’t disappear — it accumulates.