Launched this week
flatten-mcp
Resume Claude Code sessions for a fraction of the tokens
7 followers
Resume Claude Code sessions for a fraction of the tokens
7 followers
Resume your exact Claude Code session for a fraction of the tokens. flatten-mcp sets aside the files Claude already digested — the weight that rides along on every message — so you pick up the same session, lighter and far cheaper. Reversible, MIT.
Maker here. My best Claude Code sessions always got expensive right when they got good — deep into a problem, the model finally holding the whole picture.
The catch: getting there means every new message re-sends the entire history, and most of that history is files, logs, and screenshots Claude already read and drew its conclusions from. They did their job hours ago, yet they ride along on every message — so you keep paying for them, turn after turn.
flatten-mcp sets that spent weight aside: it moves the heavy tool output into a backup next to the session and leaves a small marker in its place, so the conversation resumes exactly as itself — every prompt and reply intact — just far lighter. Anything moved comes back by id the moment it's needed, and one command restores it all.
On the session that pushed me to build it: 340,071 → 132,800 tokens, 61% lighter, same conversation resumed. Fully reversible and crash-safe (atomic writes, a self-cleaning backup), Claude Code only, local, no telemetry, MIT. Happy to get into the session format or the safety model.