← All work

ReflexLoop

An autonomous agent with persistent memory, run outside any chat UI, with a console to approve or reject what it wants to do next.

Each cycle is research and a plan, then your approval, then the action, a recorded outcome, and a lesson written back into memory. Nothing with a real-world effect runs without a proposal someone explicitly approved, and the act phase is fenced to exactly the tools that proposal named. No sub-agents by design: nothing in the tool registry can spawn one.

Bring your own model. It calls OpenRouter, OpenAI, Anthropic, xAI or Moonshot directly over HTTP with no vendor SDK, and different phases can use different models. A React console shows it working and holds the approval queue.

TYPESCRIPT
REACT
NODE
SQLITE
2026