At least we've got clarity now? But a lot of my value comes from "claude -p" usage, either scheduled tasks while I'm asleep, or responding to incoming emails / voicetexts. Even the email replies will barely fit in $100/month. I'm not going to pay $1000 / month, so I guess it really is time for me to look at the competition and move my programmatic usage to them.
Man, I love the Claude models, and the whole idea of constitutional AI. We built a lot of tools & infrastructure together, but kept a lot of logs as well. I'll be really sad if I mostly have to move on now.
You need an AI for that?
When I'm away from my computer and out walking, I'll often think of a task for Claude, or I might bounce an idea back and forth with Claude via voice messages. I wrote a small Go program to watch my email and launch Claude via "claude -p" when it sees an email from myself addressed to it.
Claude also has a different "character" when collaborating over email, it feels more like a colleague. Hard to describe, but email almost feels like a better interaction UI than the chat window.
I had been starting to train Claude to see how it might go on customer service (eg maybe it could reply to my customers while I'm asleep), but at current Anthropic API costs I think that might still be too expensive.
(I could be wrong!)
I'm not using the regular email connection methods, I don't want to give Anthropic complete access to my email account. I do a ton of deterministic checks first from a Go program that actually checks each email, to avoid lethal trifecta attacks. The model technically has no access to email at all. I only give a prompt with the necessary info, and access to a custom MCP reply tool that can only email me.
Basically I'd want Cowork within my external loop, and Anthropic wants to own the loop instead. (Unless I've missed a way to do it.)
----
EDIT: Also, to the person who just tried to lethal trifecta me - nice try, but you just demonstrated all the exact reasons Cowork / claude-code needs to be within the external loop of a deterministic program. This is why you don't just dump external input straight into context, or give the model direct access to everything. We're going to see a lot more of this, not just as more people use agents, but as more hosted webmail systems decide they need to add their own homebrew AI models into everyone's systems. And seriously, German servers really need to start tightening up their security.
Maybe something similar can be done with tmux still, I'm definitely going to explore it
I use on-device STT with Claude Code's built-in remote control feature to do what you do without needing claude -p, but I guess I don't use it for large enough quantities of text where on-device STT quality becomes a big issue.
Which is interesting, since you'd also think that programmers would be their primary customers.
They are now changing it to be:
You get $20/$100/$200 of "credit" that can be used for claude -p. Problem is, once you are out of that it is the normal API rates (outrageously expensive).
I'm hoping local llms start rapidly improving even more though.
I think that this is much better than the previous situation with total lack of clarity on what is allowed and what isn't.
- The AI gives human prompts to copy-paste into Claude Code
- Human copy prompts into Claude Code
- The AI reads output from Claude Code
codex W
