A local-agent experiment, wrapped honestly.
Mercury explored a personal AI agent running through local Gemma models, Discord, a dashboard, and the Cortex bridge. The important conclusion is pragmatic: maintaining our own long-lived Hermes fork is not worth the operational cost right now. Future work should use Hermes directly and keep Red Team Kitchen additions small, documented, and upstreamable.
What We Accomplished
- Local Gemma routing: tested Gemma 4 paths on owned hardware and documented the performance/cost tradeoffs.
- Gateway surfaces: proved out terminal, Discord, and dashboard workflows around a shared local agent state.
- Cortex bridge: connected the agent stack to the brain-response pipeline and helped produce the viewer/demo artifacts.
- Negative results: documented when speculative paths and routing promises did not beat simpler local execution.
What We Are Not Claiming
Mercury is not a currently hosted public agent and should not be linked as a live service. The repo remains useful as a record of the Gemma 4 Good experiment, the scripts, the measurements, and the infrastructure lessons. The next healthy version should be a thin integration layer around Hermes, not a divergent product fork we have to babysit forever.
Repository
Source stays available at github.com/AlexiosBluffMara/mercury. New work should either happen upstream in Hermes or live in small Red Team Kitchen adapters with clear boundaries.