Most AI coding agents I tried treat my codebase like a stranger. They read a file when asked, guess at how the pieces connect, and sometimes break something they never understood.
I built Athena to work differently. It builds a real structural graph of your project first: symbols, call paths, dependencies, even dynamic-dispatch hops that grep can't follow. Then when it edits, it knows what depends on what before it touches a file.
A few things about it:
It runs in your terminal as a TUI, not a browser tab or IDE plugin. It works across providers: Anthropic, OpenAI, and more, so you're not locked into one model.
There's an /init command that generates architecture docs for your whole repo, root and per-subdirectory, so both you and the agent start with real context instead of a blank prompt. The whole stack is open source: agent loop, terminal renderer, provider abstraction, session storage. No hosted backend sits between your keystrokes and your files.
I'm one person building this because I hit the same wall over and over with existing tools. Would like feedback, especially from anyone running agents against large or legacy codebases. What breaks for you today?
Top comments (0)