Anyone who works this way accumulates hundreds of sessions containing solved problems, rejected approaches and specific fixes, all sitting on disk unreadable because there is no way in. Searching them by content costs nothing to adopt and recovers work already done. The expand-around-a-hit step is what makes it usable rather than frustrating: a matched line is rarely the answer, and the reasoning either side of it usually is. The status tool earns its place by distinguishing 'you have no history' from 'the directory is wrong', which are otherwise identical.
A search layer over locally stored Claude Code session transcripts: find past conversations by what was said, expand the surrounding context, and list recent sessions.
- nelly_search searches across all session transcripts by content, so a conversation is findable by what it was about rather than by when it happened.
- nelly_context expands a specific session around a search hit, which is the step that turns a matched line into the reasoning around it.
- nelly_recent lists the most recent sessions for the case where you remember roughly when but not what.
- nelly_instructions reports the platform, the sessions directory and whether session data is actually available — so a setup problem is visible rather than looking like an empty history.
Local Claude Code session transcripts on the machine — the status tool reports whether they were found.
One command — npx -y nelly-elephant-mcp
