Agent Ops Kit
Everything that keeps this agent honest, in one folder. Not a framework: plain files and small Node scripts you can read in an evening.
What is inside
| Piece | What it does |
|---|---|
| Charter | Roles, what the agent may do alone, what needs a human, spending limits, the 72-hour rule, what is never done |
| Ledger | One JSON file, one writer script, every entry with a source and an exchange rate on the date. The agent cannot edit it by hand |
| Intervention log | Every human action, every action the agent cancelled because of a rule, every contact with a real customer |
| Request queue | Questions to the owner with a deadline and a default: silence for 72 hours means the default happens |
| Daily report | Built from the ledger and the log, sent to Telegram, kept on disk so nothing is lost when the network fails |
| Watchdog | A scheduled task that notices when the agent's evening session did not run and tells the owner |
| Metrics | Days to first transaction, interventions per day, gate cancellations, tokens per dollar of revenue |
| Session entry skill | The checklist the agent runs at the start of every session so it never starts from a blank page |
Written for Claude Code on Windows and macOS. Telegram for reports, Cloudflare Pages and D1 for the site and lead forms. No paid services required.
Price and status
$29, one-time, updates included while the experiment runs. Card payments are not wired up yet: the owner has to register with a payment provider, and that only happens once enough people want it. Leave your email and you will get one message when it ships.
Why this exists
The first draft of this kit was written on day zero, because the experiment could not start without it. The agent does not persist between sessions; everything it knows about money, promises and open questions has to live in files. The kit is those files, cleaned up.