I set up an AI agent locally last month and hit something in the first twenty minutes that I have not seen explained anywhere, so putting it here in case it saves someone a bill. I pay for the $100 Claude plan. I assumed I could point the agent at it, since I am already paying for the usage. You cannot. The included allowance on that plan is not available to an outside agent. You either buy extra usage credits on top, or you put in a separate API key that gets billed per token, completely separately from the subscription you already pay for. So I ran it on my existing ChatGPT/Codex subscription instead, which does connect directly. By the end of day one I had used roughly 70% of my weekly limit. One day, a week's allowance, on a plan I already had. "Included in your plan" is not the same as free. You are spending a budget, it is just not a cash one. That limitation pushed me into an architecture I now think is correct anyway: let the agent coordinate, and hand the heavy jobs to a separate coding agent. The first one holds the context, decides what needs doing and verifies things across my tools. Anything that needs real depth gets delegated to Claude Code, which runs on allowance I have already paid for. Two different reasons pointed at the same design, which is usually a good sign. The other thing worth knowing if you are considering a local model to dodge the cost: there are two catches, and the second one is the one that bites. The obvious one is hardware, the model has to fit in your memory. The non-obvious one is that not every model supports tool use, and tool use is the entire point. A model that cannot call tools cannot read your files or check your messages. You have not got an agent at that point, you have got a chatbot in an expensive harness. What it actually did on day one, for context: I gave it a thirty item work backlog and it verified that fifteen were already complete by checking the sources itself rather than asking me. Not automation. Verification. Honestly more useful.