Not who you asked but I slapped this together in 100 lines of code and you may find it useful. It's just `claude -p proompt` (or indeed, `codex exec prooompt` inside a Telegram bot. (Was annoyed by NanoClaw's claim that it was 500 lines, so tried my own hand at it ;)

No memory, no cron/heartbeat, context mgmt is just "new chat", but enough to get you started.

Note: no sandboxing etc, I run this as unprivileged linux user. So it can blow up its homedir, but not mine. Ideally, I'd run it on a separate machine. (My hottest take here is "give it root on a $3 VPS, reset if it blows up" ;)

https://github.com/a-n-d-a-i/ULTRON

You may also enjoy CLIProxyAPI, which does the same thing (claude -p / codex exec) but shoves a OpenAI compatible API around it. Note: this probably violates every AI company's ToS (since it turns the precious subsidized subscription tokens into a generic API). OpenAI seems to tolerate such violations, for now, because they care about good. Anthropic and Google do not.

(Though Anthropic may auto-detect and bill it as extra usage; see elsewhere in this thread. Situation is very confusing right now.)

https://github.com/router-for-me/CLIProxyAPI

I used https://github.com/Kevjade/migrate-openclaw, and then started running Claude Code with remote exec against an empty folder that I've advised it to start adding new memories into. So far, my bot's personality is back, and it can utilize the same skills as before, which is was failing on last week.

I don't have an especially heavyweight implementation, because I only use mine to review things I've written in my Apple Notes (journaling of various kinds, mostly) and give insights.