Quickstart

This is the five-minute tour. It assumes you've installed claudex and run claude locally at least once so ~/.claude/projects/ has content.

1. See the dashboard

claudex summary

You'll get one screenful: total sessions, today/this-week counts, cost totals, top projects, top tools, model distribution, and the most recent session.

First run indexes the whole ~/.claude/projects/ tree — the spinner shows progress on stderr. Subsequent runs reuse the index (5-minute staleness window), so they're near-instant.

2. Cost, by project

claudex cost --limit 10

Aggregates token usage across every session in each project, applies the correct pricing tier per message (Opus / Sonnet / Haiku), and sorts by cost descending.

Add --per-session to break it out by individual session, or --project utensils to filter to projects whose decoded path contains utensils.

3. Full-text search

claudex search "schema migration"
claudex search serde --project claudex --limit 5

Uses SQLite's FTS5 virtual table. Case-insensitive by default; pass --case-sensitive to drop back to a file scan (FTS5 doesn't support case-sensitive queries).

Each hit shows the project, session ID prefix, date, role, and the matching line with the query highlighted.

4. Tool and model usage

claudex tools --limit 15     # How often you reach for Bash, Edit, Read, ...
claudex models               # Calls, token usage, cost per model

5. Export a session

# Find a recent session and grab its prefix
claudex sessions --limit 5

# Export as Markdown (default)
claudex export e1a2f4 --output my-session.md

# Or as JSON
claudex export e1a2f4 --format json --output my-session.json

The selector matches by session-ID prefix or by project name (substring). Combine --project for disambiguation when a short prefix is ambiguous.

Bonus: pipe to jq

Every report supports --json:

# Top three cost projects, numeric only
claudex cost --json --limit 3 \
  | jq '.[] | {project, cost_usd}'

# p95 turn duration for one project
claudex turns --project claudex --json \
  | jq '.[0].p95_duration_ms'

Bonus: live tail

When you want structured output from a running Claude Code session:

# In one shell
claudex watch

# In another shell
claude --debug-file ~/.claudex/debug/latest.log

claudex watches the debug log, formats tool calls, and inserts a banner every time claude starts a new session (which truncates the log).

See Watch mode for details.

Where next?

• The index — when it syncs, when to force a rebuild.
• Commands overview — every subcommand.
• Recipes — one-liners for common questions.