carafe: continuity layer for ai coding agents

I needed a way to hand off between coding agents, or even between sessions, when I max out & don’t want to rely on compaction. So I built Carafe.

Compaction keeps a session going by summarizing it, but you're always losing context somehow.

Carafe keeps the full record of every Claude Code and Codex session on your machine. When you hand off, it builds the handoff from that record, not from what the agent still remembers. The next agent gets an overview of the whole session, the latest work word for word, and can look up any earlier moment exactly as it happened. It also sees the current state of your code.

I’ve used it on a session that ran for five months, over 30 million tokens of history. The handoff took about 20 seconds, and the next agent picked up the work with its history intact.

The handoff is a plain file, so the next agent can be Claude Code, Codex, or anything that reads an md file. It’s open source and everything stays on your machine. And you can run it via CLI, a menu bar app and as a skill.