I make AI tools simpler.
Takes toys seriously. Takes seriousness as a toy.
I'm a product manager (VP Product). Most of my job is noticing where people get stuck and taking that step away. On GitHub I do the same thing for open-source AI agents: small fixes, plain error messages, and the odd "we could just delete this".
- Personal AI agents that run on their own computer and talk to you in chat: OpenMausBot, rakazo, airi, nanobot.
- Agent memory: claude-mem, gbrain, hindsight.
- The boring bits that decide whether people trust an agent: does it say why it failed, does it stop asking the same question, does it do what the button says.
- rakazo #1468: warn when a model can't see images, before you send one to it.
- rakazo #1461: keep a paragraph break between turns, so replies stop running into each other.
- OpenMausBot #3026: name the HTTP status when an MCP server says no, instead of a vague error.
- claude-mem #4855: switch back to the main model once it recovers.
- airi #3013: make the speech cards open the settings they point to.
- unifi-mcp #883: don't report a saved setting as failed when only the check failed.
I also write product notes on issues. Some of them ended up on hindsight's plan, which made my week.
- Find the step where people get stuck.
- Make the smallest change that removes it.
- Write the PR so anyone can review it in two minutes.
- stuckcheck: reads an AI agent's chat log and points at where it went in circles. No LLM needed.
Rule of thumb: if it needs a long README, it probably needs a shorter product.