Yerb

The Problem

Every AI demo I saw was the same thing: a chatbot. You type a question, it types an answer, and that’s it. Cool for five minutes, useless for real life.

I didn’t want an AI I could talk to. I wanted an AI that could do things. Check my calendar, submit my homework, scrape Craigslist for 4Runner deals, draft tweets, commit code — all without me asking.

The Approach

Yerb is a 24/7 autonomous AI assistant running on a Mac Mini in my apartment. It’s built on OpenClaw, which provides the gateway layer between Claude and a full suite of tools: Calendar, Canvas API (BYU’s LMS), GitHub, Telegram, browser automation, file system access, and more.

It’s not a chatbot. It’s an operating system for my life.

What It Actually Does

Calendar management. Yerb reads my calendar, sends me morning briefs, and alerts me before meetings. It knows my schedule better than I do.

Coursework automation. It connects to BYU’s Canvas API, checks for upcoming assignments, reads rubrics, and helps me stay on top of deadlines. It’s pulled assignment details, submission requirements, and grades — all programmatically.

Nightly builds. While I sleep, Yerb runs builds on my projects, checks for errors, and commits fixes. I wake up to a summary of what happened overnight.

Content ghostwriting. It drafts X (Twitter) posts based on what I’m working on, matching my voice and tone. I review and post — or sometimes it posts autonomously.

Vehicle deal hunting. Yerb runs my marketplace scraper, filters results, scores deals, and sends me Telegram alerts when something good shows up.

Knowledge management. It manages my Obsidian vault — organizing notes, generating summaries, and connecting ideas across documents.

Architecture

Me (Telegram/Web) → OpenClaw Gateway → Claude → Tools
                                                  ├── Calendar API
                                                  ├── Canvas API (BYU)
                                                  ├── GitHub
                                                  ├── Browser Automation
                                                  ├── File System
                                                  ├── Shell Commands
                                                  └── Telegram Bot

The OpenClaw gateway handles session management, tool orchestration, and persistent memory. Claude is the reasoning layer. The tools are the hands.

The key insight is that this isn’t a single script — it’s an always-on agent with context about my life. It remembers conversations from yesterday. It knows what projects I’m working on. It has opinions about my schedule.

The Key Insight

My AI assistant isn’t a demo. It’s my actual workflow.

Most people building with AI are making toys. Proof of concepts. Things that work in a demo but fall apart in daily use. Yerb has been running for months. It’s battle-tested against the messiness of real life — schedule changes, API rate limits, edge cases in assignment parsing, and the fact that sometimes I just don’t want to be productive.

The difference between a demo and a tool is reliability. Yerb handles errors gracefully, retries failed API calls, and tells me when something breaks instead of silently failing.

Stack

• Runtime: OpenClaw (gateway + agent framework)
• AI: Anthropic Claude (Opus/Sonnet)
• Language: Python, Shell
• APIs: Google Calendar, Canvas LMS, GitHub, Telegram, OpenFoodFacts
• Infrastructure: Mac Mini (always-on), macOS automation, launchd