This article is a living document - captures a snapshot of my current toolset, and updates on why I changed it.
- AI chat - huggingchat
- Chat LLM - GLM-5.2 and Nvidia Nemotron 3 Ultra (both are incredible)
- Vision LLM - Kimi K2.6
- Coding AI harness: Pi
- Hosted AI provider: openrouter
- Local AI provider: ollama
- Coding AI model:
- No screenshots needed: GLM-5.2 on openrouter
- Vision: Minimax-M3 on openrouter
- Local models (vision, thinking and tool calling):
- Qwen3.6
- Gemma4
- NotebookLM - For deep research and different “views”
- Grok - For the pulse of the internet AI chats (with X data) and also for a frank uncensored response on many things.
- Gemini - For web index data and Google properties
- Hermes agent and OpenClaw, but always in a VM - High agency agents
- OS - Mac
- Mac VM in Mac - Lume (from cua)
- Terminal - iTerm2
- IDE - VSCodium (with telemetry off, basedpyright)
- Python - uv, pydantic v2, FastAPI
- data - Files - json, markdown or DB - sqlite
- Native app - Pyside 6 for Qt
- Web frontend - vanilla html/css/js
Past luminaries
- Claude - As of today (2026-06-19), Claude and Claude Code are a mess - since Opus 4.8, the entire experience with Claude and Claude Code went south, then there was the Fable debacle, and even before that the “no subscription access, even at 200$ per month to best model” - the sota model edge is already being held by a sliver. I think they just gave it away with all the recent choices - both technical and business.
- ChatGPT - Was a favorite for a long time. They went off alignment several months back - heard they are better now, am checking them out, but no reason really to go back - cheaper models are as good if not better for most use cases.