Notes from the gate

Running AI agents against real repos, on self-hosted hardware, behind a push gateway. Field notes and honest numbers from building nimblegate.

• July 7, 2026

  Running a local AI model smoothly on a $214 used office machine

  A 35 W i5-10400T with no GPU runs a 35B-parameter model at ~8 tokens/s, with web search, RAG, and a queryable git gateway over MCP. The numbers, honestly.

Follow along via RSS or the email list on the home page.