A walkthrough of running the Codex desktop app (OpenAI's coding agent) with a local LLM instead of cloud AI, using oMLX as the inference backend instead of Ollama. The setup involves modifying a config file to point Codex at a locally running oMLX server with an MTP-enabled model (Qwen 3.6 35B A3B), which supports multi-token prediction for 1.5–2x faster decoding. Demonstrated use cases include daily news briefing aggregation from multiple sources, calendar appointment checking via MCP server skills, and general automation workflows — all running entirely on-device.
•8m watch time
122 Impressions