ArticleOctober 11, 2025Free to read

The AI and Productivity Tools on My Mac

A firsthand Mac setup using ChatWise, OpenRouter, Ollama, Cursor, and Obsidian, alongside small utilities for input, windows, search, and app switching.

Originally published . English translation: . Read the Chinese original.

Original video in Chinese.

Key Takeaways

  • I chose a 14-inch M4 Pro MacBook Pro for mobility, with plans at the time to use a Mac Studio for heavier work.
  • ChatWise supports hosted models and local tools such as Ollama, along with search and Artifacts. OpenRouter simplifies provider access through one API.
  • Cursor serves development and other AI-assisted workflows. Obsidian with Copilot supports conversations about my notes and knowledge base.
  • For personal use, I prefer simple, lightweight tools and carefully prepared data.
  • Input Source Pro, Rectangle, Raycast, and Manico improve input, window management, search, and app switching.

I had just bought a new MacBook Pro and wanted to share the software I installed, especially the AI tools that improve my everyday efficiency and experience.

Why I chose this MacBook

My configuration was an M4 Pro, 48GB of unified memory, and a 1TB SSD. I paid extra for the nano-texture display to make filming the screen easier.

Why not a 128GB M4 Max? First, the 16-inch model felt too large for carrying around; 14 inches was right for me.

Second, at the time I planned to buy a Mac Studio later that year. Industry observers expected a mid-year or later release with an M4 Ultra. I hoped it might offer substantially more memory than the MacBook, similar to the 192GB option on the M2 Ultra, and be useful for local models.

That was my plan: leave heavy workloads to a Mac Studio and use the MacBook for mobility. M4 Pro with 48GB was already powerful enough for my needs.

My footage was 4K, 10-bit, 4:2:2, edited in DaVinci Resolve. This machine handled it easily. In my local-model tests, 14B reached 21 tokens per second and 32B reached 11. I was satisfied with that performance.

Hello everyone, and welcome. I aim to explain both the why and the how of AI. At the time of this episode, newtype was approaching 1,000 paying members. Join us if you want to connect.

ChatWise, OpenRouter, and local models

After using official apps such as ChatGPT and Claude for a while, you may want API access and local open models. That lets you try more models, pay by token usage, and choose tools that fit your habits.

An app like ChatWise provides that interface. It supports major closed-model vendors and inference platforms.

I recommended OpenRouter for stable access instead of registering with numerous separate platforms. One API key lets you choose from its model selection without opening accounts and adding a credit card at every provider.

For local models, ChatWise supports Ollama and LM Studio. With Ollama running, it lists the installed models automatically. My most-used local models were DeepSeek-R1 14B and Qwen2.5 14B.

Models also need tools. ChatWise offers Artifacts and search, much as ChatGPT adds tools around its underlying model.

For search, you can use Google’s free API or Tavily, a search engine designed for models that I had used in my own scripts. Alternatively, Perplexity’s API combines search and reasoning in a model interface that I found effective.

Advanced ChatWise features required payment. Cherry Studio was an alternative for free use. I paid for ChatWise mainly because it looked good and felt pleasant to use. When functions are similar, design and smoothness matter.

I used Gemini 2.0 and Grok 3 on their official websites. For my other models, I used ChatWise with OpenRouter and Ollama.

Cursor and Obsidian

I considered Cursor the strongest overall AI programming app at that time, despite different tools having advantages in particular niches. It supported Claude 3.7 on release day. In that episode, I described 3.7 as 20% stronger than 3.5, making Cursor more useful to me. I described less conventional uses in an earlier AI learning-method video.

My other tool was Obsidian, especially its Copilot plugin. It served two purposes.

First, I could talk to AI while writing notes without switching applications and breaking my train of thought. Through OpenRouter, Copilot could use a Perplexity model for current information.

Second, it turned my notes into an AI knowledge base. The model could refer to earlier work, including notes I had forgotten, and sometimes surface useful surprises.

After trying many RAG and knowledge-base tools, I concluded that preparing the data matters more than choosing a complicated tool.

My first knowledge base contained many reports and papers. But without my own understanding and transformation, those were public materials rather than my knowledge.

I removed material that had not become part of my own understanding and kept the notes I had written myself. These are worth accumulating over time. Their simple Markdown format is easy for models to understand, so basic retrieval can be enough.

For personal AI tools, I recommend keeping the setup simple and lightweight. Heavy systems make more sense when enterprise needs require them.

Small utilities that improve the experience

Input Source Pro shows the active input method near the field you are using. When entering a URL, I sometimes typed several characters before realizing the keyboard was in Chinese mode, then had to delete and start again. The utility shows the input mode as soon as I select the field.

Rectangle positions windows through shortcuts. Control–Option with arrow keys moves a window to a half-screen position, while Control–Option–Return makes it fill the screen. It supports thirds and other layouts, though I mostly used those basic shortcuts.

Raycast helps with Google searches and local files. I invoked it with Option–Space and entered a keyword for Google. My browser address bar was set to Perplexity, giving each search tool its own convenient route.

Manico assigns shortcuts to applications. My shortcuts included Option–A for WeChat and Option–Q for Chrome, arranged around the WASD area. That habit came from playing a fire mage in World of Warcraft PvP, where quick controls matter.

Other utilities I suggested were MonitorControl for multiple displays, OrbStack for Docker, and Screen Studio for frequent screen recording.

This is one reason I like the Mac so much. With some attention to setup, it becomes a productivity tool that fits your hand and gives you a sense of control.

That is all for this episode. Join the newtype community to discuss AI. See you next time.