ctx

Working across several worktrees at once, each with its own agent session, makes switching expensive. You come back to a tab and can’t remember what the agent was doing, so you burn tokens asking it to summarize itself. There’s no single view of what’s happening everywhere. And a fresh session re-explores state that git and the transcripts already know.

ctx reads what already exists rather than asking you to maintain anything. Transcripts and git are the truth; the state directory is a disposable cache you can delete at any time and get rebuilt on the next run. One row per worktree: branch, whether a session is live, when it was last active, a distilled one line task, and whether the tree is dirty or ahead.

The same data comes out three ways. A board for scanning, a TUI for watching it live, and an MCP server so an agent can ask directly instead of re-exploring. Session hooks feed it: one records lifecycle events, another injects a recap when a session starts.

The distiller needs an LLM, but never a key. It tries the Claude CLI you’re already logged into first, falls back to an API key if there is one, and falls back again to a plain heuristic extract with no model at all. Hooks always exit zero, because a hook that breaks your agent session is worse than one that quietly does nothing.