Setting it up was easy enough, but just as I was about to start linking it to some test accounts, I noticed I already had blown through about $5 of Claude tokens in half an hour, and deleted the VPS immediately.
Then today I saw this follow up: https://mastodon.macstories.net/@viticci/115968901926545907 - the author blew through $560 of tokens in a weekend of playing with it.
If you want to run this full time to organise your mailbox and your agenda, it's probably cheaper to hire a real human personal assistant.
Not doing so feels like asking for trouble.
I'd find it hard to write such an article about how this is the next best thing since sliced bread without mentioning it spending so much money.
I load $20 at a time and wait for it to break and add more.
There has been some work around this practically being tried out using it for structured data outputs from LLMs https://docs.boundaryml.com/guide/baml-advanced/prompt-optim...
I won't claim I understand its implementation very well but it seems like the only approach to have a GOFAI style thing where the agent can ask for human help if it blows through a budget
But I was inspired to use Claude Code to create my own personal assistant. It was shocking to see CC bang out an MVP in one Plan execution. I've been iterating it all week, but I've had it be careful with token usage. It defaults to Haiku (more than enough for things like email categorization), properly uses prompt caching, and has a focused set of tools to avoid bloating the context window. The cost is under $1 per check-in, which I'm okay with.
Now I get a morning and afternoon check-in about outstanding items, and my Inbox is clear. I can see this changing my relationship to email completely.
I guess if you're letting it vibe code huge chunks. I'm doing mostly handwritten code for my current project with a little bit of "I don't want to deal with this, Claude can handle it" and I've spent $1.26 this month for my 446 lines of code.
But yes I suppose at that rate, if Gastown or Beads or whatever is 300,000 lines of code (just to use a project known to be fully vibe coded with rough LOC reported), that would be over $800.
Don't let it vibe code hundreds of thousands of lines of code I guess.
BTW, OpenCode has free Kimi (I haven't hit a quota yet) right now and it's done pretty great things for me in the last 24 hours.
My entire process is to build a generic llm.md file that all the tools can use and record to. I don't want to be tied completely to any one solution. You can get pretty far without spending a lot on tokens. I can run almost continually, and presently I'm the bottleneck anyway.
It's a lot like managing two experienced mid- to sr- engineers each of whom have slightly different personalities and intro/extro verted personalities. CC has more personality but OC wants to race. They can both code, but for disparate tasks you might pick the personality and posture of one person over the other.
I find myself picking daily tasks based on which of the tools I'm in the mood to sit with. But across a few days I sit with all three.
Even if I had to reload manually very often, I still would not enable auto reload. These APIs are crazy expensive and I'm not looking for a surprise bill.
I still have Opus review the shit out of & plan my work. But it doesn't need to be hands on keyboard doing the work.
(keep in mind with the cost savings: do an initial calculation of your cloud cost first with a low-cost cloud model, not the default ones, and then multiply times 1-2 years, compare that cost to the cost of a local machine + power bill. don't just buy hardware because you think it's cheaper; cloud models are generally cost effective)
Surely there is also the benefit of data privacy and not having a private company creating yet another ad profile of me to sell later on?
Developers trust lobsters more than humans.
The other wild thing is that many of these expensive automations that are being celebrated on X can already be done by voice using Siri, Google, or any MCP client.