Anthropic is Scrooge McDuck
Having driven OpenClaw for a few months, I decided to give Hermes Agent a spin recently because I wanted to see how far I could push it with claude as the backend.
And honestly, the setup was kinda neat.
Hermes lets you wire up different models and toolchains, and one of the options is using Claude Code CLI through ACP. On paper, that sounds great. If Claude Code is already the thing I like using directly, then being able to slot it underneath another agent feels like the best of both worlds -- Claude's coding ergonomics, but inside a more flexible agent framework.
So I went down the rabbithole.
I spent basically a whole day getting things set up. Tuning config, checking auth, wiring the backend, making sure the right tools were exposed, trying to get the whole thing into a shape where I could actually use it for real work instead of just demoing it for five minutes.
And then I got smacked with a 400.
Not a weird bug. Not some obscure setup mistake. Just an error telling me I'd exhausted extra usage.
That was the moment the whole thing stopped feeling neat and started feeling absurd.
Wait, so what am I paying for here?
Apparently, Anthropic bills extra usage for third-party ACP-based usage of Claude Code CLI.
Which, sure, maybe that makes sense from their side of the ledger. If another tool is driving Claude Code programmatically, maybe they don't want that to fall under the same mental bucket as me poking at it manually in a terminal.
But for me personally, that's where the value proposition starts collapsing.
If using Claude Code CLI as the backend for an agent still ends up charging me by usage anyway -- basically like I had just used the Anthropic API directly -- then what's the point of routing it through Claude Code CLI in the first place?
Why would Anthropic tell you that you can even do that, at all?
Why not just not force everyone to use the Anthropic API directly?
Claude Code CLI, in this setup, stops being this nice higher-level interface and starts feeling like a weird middleman. And not even a useful middleman -- just one that adds ambiguity and another place for billing semantics to get confusing.
The flip-flop is the weirdest part
What makes this whole thing extra bizarre is how much the stance seems to have flipped around.
First it was basically: no, you can't use Claude Code CLI for agents.
Then it became: yes, actually, you can use Claude Code CLI for agents.
And now the practical version seems to be: yes, you can still use it for agents, but you'll still pay usage for that flow as though you had just used the API directly anyway.
So... okay?
If that's the final shape of it, then it feels like the earlier restriction got relaxed without actually making the workflow meaningfully better for users.
Technically allowed is not the same thing as economically sensible.
And I think that's what annoys me most here. It's not even that I object to paying for usage on principle. APIs cost money. Inference costs money. Fine.
It's that the product framing suggests there's some special usefulness in being able to plug Claude Code CLI into agentic workflows, but the billing reality makes that usefulness feel mostly cosmetic -- Anthropic is just sleazily letting you think you can use your subscription and then slap you with the extra usage gotcha.
So sure, you get the complexity of a CLI-driven agent backend, but you don't really get a corresponding financial upside with the plan you're already paying for.
So what's the point?
Maybe there is a point for some people.
I genuinely thought the appeal of using Claude Code CLI as a backend for another agent would be that it gives you some kind of practical leverage -- better UX, better cost profile, better reuse of an existing subscription, something.
Instead it feels like being told you can absolutely swim in the vault now, but Scrooge still charges by the coin you touch.
And at that point, man, what's the point?
Is it too much to ask that they be like Codex and just let external usage fill the same bucket as Codex in your Pro plan?