I use and build AI agents every day, and the tools that let them do real things.
GTM Engineer by day. Nights and weekends I ship what I want to exist: agents, plugins, MCP servers, small hardware for agents, and a Claude Code OS that runs alongside my fleet of Hermes & OpenClaw agents. Everything here runs against something I actually use.
Hardware for agents
| What it does | Try it | |
|---|---|---|
| Agent Charm | A pocket ESP32 with a living character on it. Hold to talk to your agent, glance at what's waiting, hold to say yes. | No board? Run the simulator |
| Clawdito | A little desk crab that shows your Claude Code usage. | ESP32-S3 + one Python file |
| Ojo de Gato | A private pet cam from an old Mac. Apple's on-device recognition checks every motion, and your agent tells you how the cats are. | Any Mac with a camera · OpenClaw, Hermes or any MCP agent |
Plugins & tools
| What it does | Try it | |
|---|---|---|
| claude-remote-kit | Drive Claude Code from your phone, in the same session as your desk. | /remote-setup |
| Rappi | Order food, groceries, pharmacy in Colombia + Mexico by talking to your agent. | /plugin marketplace add garavitgabriel/rappi-plugin-claude-openclaw |
| ESPN Fantasy | Run an entire H2H Categories fantasy baseball season as a conversation. 44 tools. | /plugin marketplace add garavitgabriel/espn-fantasy-claude-openclaw |
| Clay Backend | Clay reads one row at a time. This lets Claude see patterns across all of them. | /plugin marketplace add garavitgabriel/clay-backend-plugin |
- Agent Charm, off the desk. Same characters, somewhere you'll always have them.
- Switchpoint: tests whether a classifier model like Jev can beat a frontier model on your task, and by how much in accuracy and cost, checked on held-out data.



