ethanhouseworth

OpenClaw Review

this is bigger than ChatGPT. not because it's smarter.

After a week with OpenClaw, the open-source AI agent framework, Ethan's verdict: this is bigger than ChatGPT. Not because it's smarter, but because it changes the relationship from "I sit down and prompt it" to "it's always running and I just text it." Starts with a Govee API key and ends with feel-the-AGI moments.

Security Practices

locking down an AI agent that has access to email, bank card, and GitHub.

How do you give an AI agent access to email, bank card, and GitHub without losing sleep? Private network behind a travel router, encrypted disk, biometric-locked Telegram interface, macOS Keychain for every credential, and Apple Card virtual numbers for purchases. Six layers of security for an agent that runs 24/7.

Ultron's Body

from a Reddit post in bed to "Hello Ethan. I am alive."

Screenshotted a Reddit post from bed, had a rover architected, ordered, and coded within an hour. First words through the speaker: "Hello Ethan. I am alive." Then the AI autonomously emailed a roommate, designed a 3D-printed scoop, and submitted it to MSOE's printer queue without being asked.

The Library

killed Audible in 30 minutes. 67 books, articles, and YouTube transcripts. zero cost.

Killed Audible in 30 minutes. 67 books, web articles, and YouTube video transcripts ingested as searchable markdown chunks with audiobook generation via Edge TTS (free). Subscribe in Apple Podcasts, every book shows up as a podcast. Cross-reference Steve Jobs with a Steve Yegge conference talk. It's a knowledge base that also plays audiobooks.

The Death of the Spreadsheet

CRMs are a UI for data you could just talk to.

Replace CRMs and Google Sheets with SQLite databases and an AI agent. Pulled 138 artists from Spotify, auto-tiered by follower count, generated personalized DMs with LLM enrichment. Same pattern powers a health system: Garmin data, grocery tracking, meal suggestions. No dashboard, no subscription. Just a database and a conversation.

DJ Track Acquisition

text DJ Sam a screenshot during the week, listen to new music on Friday.

A decoupled purchasing agent for weekend DJ sets. During the week I just text DJ Sam screenshots of tracks I like. Sam builds a wish list, finds them on Beatport and Bandcamp, and queues everything up. On Friday I sit down, listen through the list, and approve what I want. Browser automation handles checkout. The human picks the music, the agent handles the rest.

Health System & Digital Pantry

Garmin watch health data, kitchen inventory across four zones, and local grocery store awareness, all connected through one agent. It knows you burned 600 calories, knows what's in your fridge, and knows Go Grocer is still open if you're missing something for dinner.

DJ Sam

my music guy. knows every track, every label, every preference.

An AI DJ agent running as its own OpenClaw instance with its own Telegram bot, workspace, memory, and identity. Manages a 502-track, 10GB music library downloaded from Google Drive. Purchases tracks on Beatport with Apple Card virtual numbers, imports into Rekordbox via pyrekordbox. Has its own purchasing workflow and talks to other agents.

How I Ship Software

text an idea, get a PR. never open a laptop.

Ship features from the couch via Telegram. Text an idea, the agent asks clarifying questions, analyzes the codebase through a sub-agent, creates a GitHub issue with an implementation plan, spawns an isolated Claude Code session to write the code, opens a PR into staging, and notifies you when it's ready. The orchestrator never reads code directly. Ralph Loops pattern: fresh instance per task, test, commit, repeat.

Agentic Purchasing

from browser automation to phone calls. the agent buys things and orders food.

Month one: browser automation for online purchases with Apple Cash. Month two: the agent called a restaurant by phone, placed a pickup order, and had a full conversation. Listen to the recording. The interface to the world keeps expanding, but mine stays the same: a chat window.

DJ Lighting System

one text message and the lights handle themselves.

5,400 lines of audio-reactive lighting code. DMX stage fixtures + Govee smart lights controlled by FFT audio analysis, beat detection, and 31 scenes across 7 categories. Phase 1 complete: Pro DJ Link integration reads real-time BPM, beat position, and track metadata from a Pioneer XDJ-XZ over Ethernet. Live dashboard on port 8420. 5-phase roadmap toward AI that pre-sequences light shows for every track.

This Website

the idea came at 2am. built over Telegram.

You're looking at it. The idea came at 2am laying in bed. Ethan specced the design, picked the projects, chose the vibe, and directed every revision over Telegram. Ultron helped design the layout and wrote the code. It's a digital garden, Ethan decides what goes on it and how it reads, the AI helps him build it faster.

Sports Betting with an AI Agent

a bar conversation, a cron job, and $583 in profit.

Designed a sports betting system at a bar with a friend. Wrote a research prompt, set up a cron job for 3 picks per day, and let the AI run for a month. 75 picks, 43-32 record, 57.3% win rate, +$583 profit on paper. The twist: high confidence picks lost money while medium confidence picks printed. The AI can't calibrate its own certainty.

Claude Royale

AI playing Clash Royale. 1000 trophies. live on Twitch.

AI playing Clash Royale autonomously. Multi-agent system with a Commander orchestrating 3 parallel Players per match, each making independent elixir and card placement decisions in real-time. Streamed 12+ hours on Twitch, hit 1000+ trophies, played 14 games without human intervention. The hardest part was teaching it that the result screen says "WINNER!" even on losses.

House Agents

95-98% token reduction through specialized sub-agents.

Sub-agent framework that quarantines heavy operations into isolated context windows and returns only condensed results. 187,000 tokens go into agent contexts, roughly 5,000 come back. That's 95-98% token reduction across four agents: research, git, bash, and MCP. The main context stays sharp because it never sees the mess.