Skip to main content

​
What is Turnlog

Your coding agents write a transcript of everything they do. Claude Code, Codex and Cursor all keep one, on your disk, in your home directory. Between them they hold every prompt you wrote, every command that ran, every file that changed and every error that came back — and none of them give you a way to look through it. Turnlog is that way. One command, no account:
It indexes what is already on your disk into SQLite, opens a local web UI, and lets you search and replay the lot.

​
What it is for

Finding the thing you already solved. You fixed this a month ago, in a session you cannot name, in a repo you have since renamed. Full-text search across every agent, then like: to find the other times it came up. Reading a run back. A 5,000-message session collapses to a spine of prompts, each with a mechanical summary of what happened under it. Lenses cut it down to just the diffs, just the commands, or just the errors. Knowing what it cost. Cost by day, model, or project — and, uniquely, cost filtered by a search query, because “what did this kind of work cost me” is not a question a dashboard can answer without a content index. Giving your agent a memory. turnlog mcp serves the same index to any MCP-capable agent, read-only, so it can search its own past work mid-task.

​
What it is not

It is not a cost dashboard with a log viewer bolted on, and it is not a hosted service. There is no account, no seat, no cloud, and no upsell — it is MIT-licensed and free, and every version stays installable from npm forever. It also never writes to your agents’ data. ~/.claude, ~/.codex and ~/.cursor are opened read-only and never modified; Cursor’s IDE database is copied before it is read, so the original is never even opened. See Privacy.

​
Every agent, one history

Turnlog reads Claude Code, OpenAI Codex, and Cursor — both the cursor-agent CLI and the IDE’s own chats. Work you did on one repo with three different agents reads as one timeline, because a project is the repo, not the tool. Which agent wrote a session is always visible, and agent:codex narrows to one. New agents are added at the normalized layer, so every feature — search, replay, spend, export, MCP — works for a new agent the day its adapter lands. Nothing is built against one vendor’s file format.