Astra-Claw
Astra-claw is a terminal AI coding agent I built in Python. You type a request, it streams a response, and it can call tools to run shell commands, read and write files, search, and remember things across sessions. File edits go through a preview-and-approve gate, so you see a diff before anything changes on disk.
I built it to understand agents from the inside — the tool loop, streaming, session persistence, and memory — and then I spent a week actually using it on boring, real tasks to find out what holds up. I wrote about that week here: I built my own AI coding agent. Then I actually used it.
What it does
- Tool loop with visible steps — shell, file read/write/patch, and search, with intermediate actions shown
- Approve-before-edit — every file write shows a unified diff and waits for
y/n/a - Persistent memory — keeps user facts and lessons across sessions
- Markdown skills — drop-in extensions discovered from a skills folder
- Workspace fencing — restrict edits to a sandbox directory
- Model switching mid-conversation, with OpenAI and OpenRouter support
The code, and an unedited friction log from the dogfooding week, live in the GitHub repository. MIT licensed.