Track what your AI agents actually cost per task, not per month

Source: Dev.to

The problem

Your AI agent just finished a 45‑minute coding session.
It edited 15 files, made 4 commits, and called the API 23 times.

What did it cost?

No idea. You’ll find out on your next invoice aggregated with everything else.

The fix

I built codesession‑cli — a CLI that teaches your AI agent to track its own costs.

Setup (30 seconds)

npm i -g codesession-cli
clawhub install codesession

Start a new OpenClaw session. Done.
The agent reads the skill instructions and handles everything.

cs start "Fix auth bug" --close-stale --json
# → agent works on your task normally…

cs log-ai -p anthropic -m claude-sonnet-4 --prompt-tokens 8000 --completion-tokens 2000
# → logs its own token usage, cost auto‑calculated

cs end -n "Fixed bug, added tests" --json
# → Session: 9m • 3 files • 1 commit • $0.15

You don’t touch anything. Just review the data when you’re curious.

cs stats
# Total: 50 sessions • 8h 34m • $47.23 AI cost

Features

• Token spend per API call (17+ models with built‑in pricing)
• File changes and git commits (via git diff, not a watcher)
• Session duration and cost summary
• Annotations — the agent leaves breadcrumb notes as it works
• TypeScript, SQLite (WAL mode)
• Local storage (~/.codesession)
• JSON output on every command — parse schemaVersion for forward compatibility
• Structured errors: { error: { code, message } }, always exit 1
• Sessions scoped by git root, not cwd
• MIT licensed

GitHub:
npm:

Looking for early adopters

What cost queries would you want after a month of data?