Your AI pair programmers leave logs; this tool adds up the tab

What it does

Tokentab is a local CLI that parses session logs left behind by Claude Code, OpenAI Codex, and Gemini CLI to calculate exactly how much you’re spending on AI-assisted coding. It groups the damage by model, project, day, and even tries to categorize the kind of work—debugging, testing, refactoring, or just chatting. Everything stays on your machine; there are no accounts to create and no telemetry to worry about.

The interesting bit

Instead of guessing or querying live APIs, the tool relies on a hand-curated, offline price table and fuzzy model-name matching, which means a vendor renaming a model overnight won’t crash your report. It also handles vendor-specific cache accounting—subtracting Gemini’s included cache hits and tracking Claude’s separate cache reads—so you don’t pay twice on paper for tokens you already paid for once.

Key highlights

• Only third-party dependency is rich; the web dashboard, parsers, and HTTP server are pure Python standard library.
• Serves a local-only web dashboard at localhost:4747 with charts and monthly statements—no CDN fonts, works offline.
• Automatically strips color codes when piped, making it safe to paste into PRs or chat logs.
• Adding support for a new coding tool is a single module implementing a collect() function; pricing and reporting are decoupled from provider specifics.
• Exports JSON for further analysis and supports filtering by date range, provider, or project.

Caveats

• Cursor support is currently stubbed out and not yet functional.
• Model pricing is maintained manually, so newer or renamed models may show as $0.00 until the table is updated.
• Activity labels like “debugging” or “refactoring” are deterministic heuristics based on your first message, not ground truth.

Verdict

Worth a look if you’re running up a tab across multiple AI coding tools and want to know which project or model is eating your budget. If you only dip into a single chatbot occasionally, it’s probably overkill.