GoKawiilA few months ago, I stood up a fiddly daily cron-job to log into my bank, credit card, and brokerage/retirement accounts. It was powered by Codex CLI running non-interactively, and had a fairly simple job: using the Chrome DevTools MCP, log into each website, extract balances and recent transactions, then send my wife and me a sort of “daily financial overview” email. It actually worked surprisingly well.
Well, until the next day. Then it broke. This pattern continued for some time — more often than not, it would work, but then some browser rendering quirk would break that day’s run, or we’d get an unexpected 2FA prompt that would brick that account. Sometimes, GPT would decide to completely change the email format, or it would get confused during a run and only pull one account’s information. Then, there was an account we needed to add that only allowed login with passkeys.
Eventually, I became a sort of unpaid tech support person for my wife, who wasn’t pleased when the daily emails didn’t come through as expected. Somehow, this led to my building Driggsby, an MCP server that tries to make this process less stressful. Two months, 75k lines of Rust, and a Plaid contract later, here we are.
Routines as an unlock
At a high level, Driggsby hooks into financial accounts using Plaid. Then it exposes a variety of tools over MCP: balances, transactions, investment info, loan info, and the like are each accessible through their own tools.
At first, the only way I used Driggsby was interactively, and on-demand — whenever I had a financial question, I’d open Claude, give it my question, and it answered using Driggsby. Over time, patterns started to emerge: most of my queries were about net worth, reviewing balances/transactions, and monitoring investments. Which made me wonder: what would be the easiest way for me to set this up on autopilot?
A few days ago, Claude Code routines were released, and immediately piqued my interest. As a Claude Code user already, this seemed like it would make it even easier to deploy an agent that would do anything I wanted.
Now, there’s nothing new per se about an agent loop running in the cloud. There are a million of those. What does feel new with Claude Code routines is the ease of setting one up.
You don’t have to write a bunch of agent-loop code and figure out where to deploy it, or even spin up OpenClaw, the Codex SDK, or claude -p on Hetzner. You just write a prompt. If you can get your data/tools into your routine cleanly through some sort of MCP connector, you can cook.
The daily email
... continue reading