Clerk

繁體中文 | 日本語 | 한국어

Most AI tools are built for AI. clerk is built for humans.

Your Claude Code sessions disappear when you close the terminal. clerk turns them into a searchable knowledge base that you own.

The Problem

It's Friday afternoon. Time for the weekly report. You open git log and try to reconstruct what you actually did across three projects, eight sessions, and five days. Half the work isn't even in git — it was debugging, research, architecture decisions, conversations with Claude about trade-offs you've already forgotten.

Claude Code doesn't remember across sessions. And you shouldn't have to.

Why not just ask Claude?

You could try. Open Claude Code and ask "what did I do last week?"

It doesn't know. It only sees the current session. To look back, you'd need to find the right session ID, load it with --resume, ask for a summary, then repeat for every session across every project. Each time, Claude re-reads the entire raw transcript — burning tokens to reconstruct what could have been saved in a single markdown file.

clerk takes a different approach: one API call per session, at the moment it ends, producing an incremental summary stored as plain markdown. By Friday, your week is already summarized. clerk report --days 7 reads those summaries and generates a structured report in one shot.

The Solution

That's it. clerk runs entirely on your machine — no remote services, no accounts, no data leaving your laptop.
About token usage: clerk uses AI to generate summaries and reports. By default it uses Claude Code (claude -p), but you can switch to any OpenAI-compatible provider (Groq, Gemini, OpenAI, Ollama, etc.) via clerk config set summary.provider.name . Each session end triggers one API call for the summary, and each clerk report invocation is…