I text a message to a bot. A few seconds later a full coding agent — Claude Code, the actual model, not autocomplete — wakes up on hardware I own and starts working. It reads, it writes, it runs commands, it ships. I approve each step with a tap. I do this from my phone: on the couch, on a walk, at three in the morning.
In the last twelve hours it burned about seventy dollars of API.
Sit with that number, because it's the whole point. Seventy dollars. In twelve hours I watched an agent restore a fleet of downed services, close a live security hole, redeploy production, ship several thousand words of finished writing, and trace an investigation back to its primary source — most of it while I was nowhere near a desk. A contractor bills that as a week. It cost me the price of a dinner.
Here's what I actually did, technically, because the trick is dumber than it sounds. I hacked Claude Code onto Telegram. A long-polling loop, the agent SDK, and every tool call the agent wants to make — every file edit, every shell command — gets routed to my phone as a Yes / No button. It can't touch anything without my tap. It's confined to a working directory. Every token is metered. That's the whole hack.
And it changes everything, because the model was never the bottleneck. The interface was.
We've spent two years arguing about how smart these things are while quietly leaving them trapped in a browser tab you have to sit in front of. The capability was already there. What wasn't there was a way to carry it — to hand it a task from a bus, get a question back an hour later, tap yes, and find the work done. Put a capable agent behind a thread you can answer from your pocket and "AI assistant" turns into something else entirely: an operator on call. The unlock is ergonomic, not intellectual.
The economics are the part people haven't felt yet. I'm not paying a salary. I'm renting judgment by the token, and it's metered to the penny. Seventy dollars bought a day of senior work I can see the exact cost of, line by line, in a usage log. There's no per-seat license, no minimum, no HR. When it's idle it costs nothing. When I need a week of work at midnight, it's there, and the bill is a rounding error against what it replaces.
I want to be honest about the shape of it, though, because the hype version is a lie. This is not a machine that thinks for me. Every one of those actions was mine to approve or kill, and a real amount of the skill is knowing what to ask, in what order, and when to say no — when the agent is confidently about to do the wrong thing and the only thing standing between it and production is that I read the diff. It amplifies judgment. It does not supply it. Point it at a problem you don't understand and it'll help you build the wrong thing faster than you ever could alone.
But pointed by someone who knows what they want, held behind a yes, it is the most leverage I've ever had in one hand. I've spent my whole career trying to do the work of a team by myself. For the first time the math actually closes.
The reason I'm writing this down instead of just enjoying it: it's good enough that I'm turning it into something anyone can run. Bring your own model login. Tether your own git. Get a full agent — code, commit, deploy — living in a Telegram thread, gated behind your own approvals, on infrastructure you don't have to think about. A deployment stack you drive by text.
The future everyone pictures is a smarter chatbot in a window on a laptop. I don't think that's it. I think it's an operator in your pocket that you trust precisely because you hold the yes — capable enough to do the work, contained enough that it can't do it without you.
Seventy dollars and twelve hours to prove it. The tab's still open, and the agent's still waiting for the next tap.