Setting Up Personal Automation with Claude Code (Non-Financial Guide)
Claude Code is a coding agent you run from a terminal — it can read files, run scripts, and call APIs on a schedule. This guide covers the non-financial parts of setting it up as a personal automation layer: keeping a machine always on, storing API keys safely, and getting a Telegram notification when something happens. It does not include trading code, buy/sell signals, or any investment advice.
What this setup actually is
At its simplest, a personal automation setup with Claude Code is three parts: a machine that stays on, a scheduler that starts a task at a fixed time, and a notification channel that tells you the result. None of that is specific to finance — the same pattern works for checking a website, summarizing an inbox, or watching a public data feed. The finance-adjacent version people often ask about — reading a market data API and sending yourself an alert — is still just a scheduled read-and-notify job, not a trading system.
1. An always-on machine
A scheduled task only runs if the machine is on at the scheduled time. Most people use either a spare desktop or laptop left running at home, or a small always-on VPS. A local PC is free if you already own one but depends on your home power and internet; a VPS costs a monthly fee but stays reachable independent of your own connection. Either way, the job is triggered by the OS scheduler (Task Scheduler on Windows, cron or systemd timers on Linux) calling Claude Code non-interactively (headless mode) at a fixed interval.
2. Storing API keys safely
Any automation that reads a third-party API needs a key or token. The basic rules that matter most for a solo, non-financial setup:
Keep keys out of your scripts and out of git
Put keys in a separate file or environment variable that your script reads at runtime, and add that file to .gitignore before you ever commit. A key pasted directly into a script gets copied every time the script is copied, shared, or accidentally pushed to a public repository.
Use the narrowest scope and rotate on a schedule
If the provider offers read-only or scoped keys, use them instead of a full-access key — a leaked read-only key can see your data but cannot act on it. Rotate keys periodically and immediately after any exposure, such as accidentally logging one to a shared file.
Best for: anything connected to a financial or personal account3. Telegram notifications
Telegram's Bot API is a common, free way to get a one-line alert from a script: you create a bot with Telegram's BotFather, get a bot token, and your script sends a message to your own chat ID with a plain HTTP request when the scheduled job finishes. This is a one-way notification channel, not a control channel — treat any two-way bridge (where you can reply to trigger an action) as a more sensitive setup that deserves its own access controls.
4. What Claude Code actually costs
Per Anthropic's own pricing page, the Pro plan — which includes Claude Code — is $20/month billed monthly, or $17/month billed annually (charged as $200 up front). Higher-usage Max plans exist above Pro at separate prices for heavier automation workloads. Every plan has usage limits that reset on a rolling session window, with weekly limits added on paid plans. Prices and limits change, so treat the figures above as a snapshot and check the official page before you commit to a plan sized for round-the-clock scheduled jobs.
Source: claude.com/pricing (accessed September 2026).
Source: Financial Supervisory Service (금융감독원), quasi-investment-advisory business guidance — fine.fss.or.kr (official consumer portal, accessed September 2026).
5 pitfalls we hit running this exact pattern
These are honest, experience-based notes from running scheduled Claude Code jobs with a Telegram bridge ourselves — not hypotheticals.
1. A silent CLI update can break every scheduled job at once
When the command-line tool auto-updates and the update fails partway, the command can become a broken stub. Every scheduled task that calls it then fails silently — no error dialog, because nothing is watching a headless job. The fix is a small daily health check that actually runs the command and confirms it returns a real version string, not just checks that a scheduled task "exists."
2. A full disk kills headless sessions without an obvious error
If the drive running the scheduler fills up, new sessions can fail to start or write logs, and the failure often looks unrelated to disk space at first. Check free disk space as part of the same health check, not as a separate afterthought you only remember after something breaks.
3. "Sent" is not "delivered"
A script logging that it sent a Telegram message only proves the HTTP call didn't error — it does not prove the message reached you (a revoked bot token, a wrong chat ID, or Telegram-side rate limiting can all swallow it silently). If a notification is important, verify delivery with a separate periodic check, not just the sender's own optimistic log line.
4. A checker that only detects problems creates endless manual fixing
A monitoring script that flags an issue every run but never verifies the fix actually worked turns into a loop of the same alert firing again and again. Build the checker to re-test after a fix is applied, not just to report the original symptom on a timer.
5. One noisy self-test can look like real signal
If your own automation calls the same API or website it is supposed to be monitoring, its own traffic can pollute the very metric you're trying to read. Separate your monitoring calls from any calls that touch a live account or public-facing count, or you will end up debugging noise you created yourself.
Frequently asked questions
Is this guide about automated trading or trading bots?
No. It covers general personal automation with Claude Code — scheduling, API key storage, and Telegram notifications. It includes no trading code, no buy/sell signals, and is not investment advice.
How much does Claude Code cost?
Per Anthropic's official pricing page, the Pro plan (includes Claude Code) is $20/month billed monthly or $17/month billed annually. Higher-usage Max plans cost more. Confirm current pricing at claude.com/pricing.
Is it legal to connect an exchange or brokerage API for personal automation?
Reading your own account or a public price feed for a personal alert differs from advising the public. Offering individualized investment judgments to others for compensation is regulated in Korea as a 유사투자자문업 (quasi-investment-advisory business) and must be reported to the Financial Services Commission. This guide does not provide investment advice — verify your own use case against the Financial Supervisory Service's official guidance.