What is claudex?

claudex is a Rust CLI that reads the JSONL transcripts Claude Code writes under ~/.claude/projects/, ingests them into a local SQLite index at ~/.claudex/index.db, and exposes reports as subcommands.

Claude Code already persists every conversation you have with it — every user message, every assistant reply, every tool call, every token-usage block, every file edit, every PR link. Those files are flat JSONL logs, and they're hard to query by hand. claudex turns them into something you can actually ask questions of.

The shape of the tool

 ~/.claude/projects/<encoded-path>/<session>.jsonl
         │   (Claude Code writes these)
         ▼
 claudex parser  →  SQLite index (~/.claudex/index.db)
         │
         ▼
 claudex <subcommand>  →  table + palette on TTY, --json for pipelines

Every read command:

• Supports --json to emit a stable, machine-readable shape.
• Honors --color auto|always|never (and NO_COLOR) for color output.

Most Claude Code reports also:

• Use the index by default (incremental sync, 5-minute staleness window).
• Support --no-index to bypass the index and scan JSONL files directly — the fallback path always matches the indexed path.

Two caveats: models, turns, prs, and files are derived from the index only and don't accept --no-index. And claudex codex is a separate reader — it scans ~/.codex/ directly and never touches the Claude Code index. See the flag support matrix for the per-command breakdown.

Who is it for?

You already use Claude Code and want to:

• Understand where your token spend is going (per-project, per-session, per-model).
• Search across past conversations without grepping JSONL by hand.
• See which files get modified the most across all your projects.
• Track how sessions turn into PRs.
• Measure turn latency (avg, p50, p95, max).
• Export a past session as Markdown or JSON.

If you've never run claude locally, there's nothing for claudex to read — ~/.claude/projects/ will be empty.

What it isn't

• Not a launcher for Claude Code. You still start sessions with claude; claudex just reads what those sessions wrote.
• Not a sync service. Everything lives locally. No network calls.
• Not authoritative pricing. Costs are approximate — they apply published Opus / Sonnet / Haiku tiers to the token-usage blocks in each record. See Pricing model.

Where to go next

• Installation — Cargo, Nix flake, from source.
• Quickstart — the five commands you'll run in the first minute.
• How it works — parser, store, index, and how fallbacks stay in sync.
• Commands overview — every subcommand, with flags and examples.