claudex

claudexQuery your AI coding sessions

A Rust CLI that indexes the local transcripts of Claude Code, OpenAI Codex, GitHub Copilot, Pi, OMP, Gemini CLI, OpenCode, Cline, Roo Code, and OpenClaw into one SQLite database and turns them into reports — cost, tools, turns, PRs, full-text search, and more, across every provider.

Six providers, one index

Claude Code (~/.claude/projects), OpenAI Codex (~/.codex), GitHub Copilot CLI (~/.copilot), VS Code Copilot Chat, Pi (~/.pi/agent), OMP, Gemini CLI, OpenCode, and OpenClaw (~/.openclaw) are all first-class. Every report spans them by default; narrow with --provider.

Additive & retentive

Archive or delete a session from disk and its indexed data stays. Non-destructive schema migrations and per-provider incremental sync mean historical usage never disappears.

Reports out of the box

summary, sessions, cost, tools, models, turns, files, prs, search, export — plus a live log watcher and a skill generator. Read commands support --json; Claude reports also support --no-index.

Proper filtering everywhere

--provider, --project, --model, and --since / --until (dates, RFC3339, or spans like 7d / 2w) on every reporting command. --json always carries a provider key for unambiguous scripting.

Honest pricing math

Fable / Opus / Sonnet / Haiku and OpenAI Astra / Sol / Terra / Luna / gpt-5 / gpt-4 tiers, applied per model. Pi and OpenClaw report their own per-message cost when available (local models = $0). Sub-cent values fall back to four decimals.

FTS5 full-text search

The index ships with a messages_fts virtual table. Search across every user and assistant message in every session and provider, with SQLite's FTS5 ranking.

Worktree & subagent aware

Worktree sessions aggregate to the parent project; Claude subagent transcripts roll up to their parent session. Group-by-project queries do the right thing automatically.

Single binary, no daemon

Built with rusqlite (bundled), clap, and owo-colors. Runs on Linux and macOS. Nix flake included. cargo install or build straight from source.

At a glance

$ claudex summary

# How much have I spent on Codex this month?
$ claudex cost --provider codex --since 30d

# Top 5 projects by cost across all providers
$ claudex cost --limit 5

# Which files get touched most across all my projects?
$ claudex files --limit 10

# Find every session where I discussed "migrations"
$ claudex search migrations

# Export one session as Markdown
$ claudex export <session-prefix> --output session.md

Install

bash

# With Cargo
cargo install claudex-cli

# Or with Nix flakes
nix run github:utensils/claudex -- summary

# Or from a local checkout
git clone https://github.com/utensils/claudex
cd claudex && nix develop && cargo build --release -p claudex-cli --bin claudex

See the full installation guide for Nix, devshell, and shell-completion setup.

Why claudex?

Claude Code, OpenAI Codex, GitHub Copilot, Pi, OMP, Gemini CLI, OpenCode, Cline, Roo Code, and OpenClaw each persist every conversation as local transcripts — every user turn, every assistant message, every tool call, every token-usage block, every file modification — but those files are flat logs in several different formats, not a queryable store.

claudex reads them once, normalizes them into one SQLite index, and gives you a CLI that answers questions like:

How much have I spent on Codex versus Claude this month?

Which project burned the most Opus tokens last week?

What's my p95 turn duration in this repo?

Show me every session that linked a PR.

Full-text search: where did I first discuss the schema migration?

How many times have I edited src/index.rs across sessions?

The index is additive — sessions you archive or delete from disk stay in your history. No cloud. No daemon. No background service. Just a small Rust binary and a SQLite file under ~/.claudex/.