Agent Ledger — Development skill for Claude Code

Agent Ledger

**Understand your agent's work. Recover the context worth keeping.**

A local workspace for Claude Code and Codex history. Recover why decisions were made, understand session usage, and save useful context for your next session. Optionally let your coding agent retrieve that context through a read-only MCP connection.

No website, hosted service, account, model subscription, or build step. Clone the public source and run it on your computer. MIT licensed; Python 3.10+ and its standard library, including SQLite FTS5.


*Recorded from `python3 ledger.py --demo`. Everything shown is generated example data.*

Run locally

In the cloned repository directory:

python3 ledger.py

Windows: `py -3 ledger.py`. If `python3` is unavailable, use your Python 3 command. Keep the terminal running; **Ctrl+C** stops the app. It opens a local browser tab. If that fails, paste the full URL printed in the terminal, including its temporary access fragment.

The first run builds its own index; it does not reuse another memory app's cache. Large histories can take several minutes. Progress identifies Codex versus Claude Code and shows checkpoints through large files. Wait for **READY — your real history**, then use that run's printed URL. Later imports skip unchanged files. If interrupted, restart with the same options to reuse checkpoints; do not delete the cache.

To explore without reading any real history:

python3 ledger.py --demo

Demo mode uses generated Claude and Codex examples in a temporary directory, removed on normal exit. Saving notes in demo mode is temporary too.

Ask “why did we decide that?”

1. Select a project and open Ask a project. Ask about a specific decision, failed approach,