Skip to content

Quickstart ​

Run Chronicle and reach your first time-travel moment — clicking a message and watching your code snap back to how it looked at that instant — in a couple of minutes.

Chronicle's core trick is simple to experience but hard to forget: it lines up every message in an AI coding session against your Git history, so any point in the conversation becomes a window into the exact state of your code. Here's the whole loop in a few seconds:

No account, no API key, no network — Chronicle reads the logs your AI tools already wrote to disk, entirely on your machine.

1 · Launch it ​

bash
npx chronicle-cli

That's the whole install. npx fetches Chronicle, starts a local server (default http://localhost:41730), and opens your browser to the dashboard automatically. Requires Node.js 24+ — see Installation if you need to check or upgrade your Node version, or want the --port/--no-open flags. Press Ctrl-C in the terminal to stop it whenever you're done.

2 · Land on Home ​

The dashboard opens on Home: a day-grouped stream of your most recent sessions across every tool, with search up top and a project rail on the right. First run, this is empty — the next step fills it in.

3 · Import your sessions ​

  1. Click + Import Sessions (top bar, available from any view). Chronicle scans the standard log locations for all four supported tools — Claude Code, Codex, Cursor, and OpenCode — and shows what it found on your machine, no configuration required.
  2. Pick a source — Claude Code is usually the richest if you have it installed.
  3. The wizard lists sessions with NEW / Partial / Imported badges (new ones are pre-selected). Hit Start Import — it's read-only, so your original logs are never touched.

After the first import, Chronicle keeps itself in sync automatically — new sessions and updates to in-progress ones are picked up in the background as you use your AI tools, with no manual re-import needed. See Supported tools for the full tool matrix and log locations.

4 · Pick a Git-backed project ​

Chronicle time-travels through Git commits, so open a session whose project is a Git repository with some history. The more commits it has, the more precisely Chronicle can reconstruct code between messages. A project with no repo still plays back the conversation — you just won't get the code-snapshot pane.

5 · Time-travel ​

  1. From Home, click a project card on the right rail, then click any session — or open one straight from the recent-sessions stream.
  2. The session opens on Overview — stats, cost & usage, agent-active time, and a Subagents card if the session used any. Switch to Playback from the left rail (or press ⌘2, or Ctrl+2 on Windows and Linux).
  3. Click any message. The middle pane rebuilds your file tree and file contents as they were at that moment, resolved to the nearest preceding commit. Changed files are green-dotted and auto-selected. Press D for the diff.
  4. Drag along the timeline (bottom) to scrub the whole session and watch the code evolve commit by commit.

That's the "aha" — the whole point of Chronicle.

6 · Explore Insights ​

Click Insights in the sidebar for the cross-project view — three tabs: Overview (spend, tokens, agent-active time, tool mix, and a top-sessions-by-cost table across everything you've imported), Explore (a pivot table — slice spend/tokens/requests by model, project, source, tool, skill, or subagent), and Content (what's actually filling your context — tool results, skills, and subagents, as a share of total tokens). The same three tabs exist scoped to a single project (its Overview / Explore / Content / Sessions tabs) and scoped to a single session (from that session's Overview, drill into Content).

Local-first: Every step ran entirely on your machine. Chronicle made no model call and no network request — it parsed local logs into a local SQLite database and reconstructed code from your own commit history. Nothing about viewing a session leaves your laptop. The one feature that does run a model is Ask, which is off by default; turned on, it runs claude -p locally on your own Claude subscription, once per question you type.

  • Installation — the Node requirement, CLI flags, and where data lives.
  • Supported tools — the tool matrix, log locations, and configuration.
  • How it works — the ingestion pipeline, the Git snapshot engine, and the Insights engine in depth.

Released under the Apache License 2.0.