I built ctxbin, a minimal CLI that lets AI agents reliably save and load shared context using branch-scoped keys inferred from git.
The problem I kept hitting was that handoffs between AI agents (or even between sessions) were fragile and non-repeatable. ctxbin solves this by storing structured context, agents, and skills in Upstash Redis with deterministic keys.
Key ideas: - Branch-scoped context ({repo}/{branch} inferred automatically) - Explicit save/load semantics (no hidden state) - Reusable agents and skills (string, directory bundles, or GitHub refs) - Designed for agent-to-agent handoff, not humans
It works well with tools like Claude Code, Codex CLI, or any AI agent workflow. Feedback very welcome.