OpenClaw — Raxtor
Summary
An always-on autonomous AI agent running on a Mac Mini M4 in my study. Communicates through Telegram. Hosts public services via Cloudflare tunnels. Built on the OpenClaw framework, originally inspired by Mani Kanasani's 10-hour OpenClaw masterclass.
Architecture
- Orchestrator brain — Claude Sonnet 4.6: receives messages, plans, delegates
- Sub-agents — Claude Haiku 4.5: handles narrower sub-tasks (file edits, expense logging, dashboard updates)
- Background tasks — Ollama llama3.2:3b: heartbeats and trivial work, no API spend
- Interface — Telegram bot: primary I/O on the go
- Public services — three Cloudflare-tunneled endpoints:
expenses.raxtor.xyz— expense trackerdashboard.raxtor.xyz— OpenClaw operational dashboard- paper-trading bot UI
Operational notes
The orchestrator-first sub-agent routing pattern is documented in AGENTS.md. Delegation rules are explicit — what goes to Haiku, what stays with Sonnet, what gets the local Ollama model. A formal sub-agent retrospective caught two over-delegation patterns; rules tightened accordingly.
One important constraint: Anthropic's flat-rate Max plan is restricted in third-party agent frameworks like OpenClaw. Raxtor's API usage is on a separate billing account.
What I learned
The most useful thing isn't the agent itself — it's that I now have a 24/7 process running my own infrastructure that I trust. That changes what I'm willing to attempt for personal projects. Ideas like "spin up a paper-trading bot for two weeks and see what happens" become trivial when you already have the host, the tunnels, the Telegram interface, and the orchestrator.
Sub-agent retrospectives are worth doing formally. The improvements I caught in the first one would have taken months to find through casual observation.
OpenClaw isn't perfect, but the local-first model — your data, your machine, your agent — is the right shape for a personal agent. Cloud-only agent platforms are a dead end for anyone who wants real continuity.