monitor-ccu — AI Best Practices

What it is

monitor-ccu is a small local server that reads Claude Code's JSONL session files (in `~/.claude/projects/**/*.jsonl`) and exposes a web dashboard at `http://localhost:3333`. It aggregates: cost, input / output / cache tokens, model used, tool calls, latency, active sessions, sub-agents.

There's also `monitor-codex` (by the same author) that does the same job for OpenAI Codex by reading `~/.codex/sessions/**/*.jsonl` on port 3334. Both can run side by side, which lets you compare Claude and Codex on similar tasks.

How it works

No API key, no external data sent. The server polls locally on the session files written by Claude Code itself. You start `monitor-ccu`, open `localhost:3333`, and you see in near real time what your day costs.

On the UX side, the dashboard surfaces the KPIs that matter: cost per feature or per project, cache hit rate, tool-call rate, refusal rate, latency p50 / p95 / timeout, top 10 most expensive prompts. It's the equivalent of a Grafana, but for your agent.

How to install it

Clone the repo, install dependencies, run:

git clone https://github.com/tiber76/monitor-ccu
cd monitor-ccu
yarn install   # or npm install
yarn start     # server on localhost:3333

Open `http://localhost:3333` — the dashboard loads sessions present in `~/.claude/projects/`. No configuration required.

How I use it daily

I leave it running permanently in a tab. When I don't understand why one day is more expensive than another, I head there: in two clicks I can see which sub-agent blew its budget or which command spilled too many logs.

It's also the only honest way to measure the impact of an optimisation: before / after adding a more compact CLAUDE.md, before / after installing RTK, before / after depth routing. Without this dashboard, I'd be reduced to gut feelings.

Sources & credits

monitor-ccu and monitor-codex are designed and maintained by `tiber76`. Full credit for the design and implementation goes to them.