Huang Yihe's AI tool stack usage and recommendation evolution from 2024-03 to 2026-05. Covers 178 posts in the tools-stack topic bucket. Focuses on "how non-technical individuals choose tools + how to build a combined stack".
When to open:
- You want to run large models locally and don't know whether to use Ollama or LM Studio
- You want to build a knowledge base and are torn between AnythingLLM / Open WebUI / RAGFlow / Obsidian + Copilot
- You want to buy a Mac mini M4 to run local AI and want to know if it's worth it
- You want to see the evolution of Lao Huang's tool stack recommendations from 2024 to 2026 (many tools have been eliminated)
0. Evolution Timeline
graph TD
A["2024-03 ~ 06<br/>本地模型 + 知识库<br/>(Ollama / LM Studio /<br/>Open WebUI / AnythingLLM)"] --> B["2024-08 ~ 11<br/>RAGFlow / MaxKB / Perplexica<br/>更强解析 + 本地搜索引擎"]
B --> C["2024-12 ~ 2025-02<br/>DeepSeek-R1 风暴<br/>Mac mini 本地服务器<br/>Cherry Studio / ChatWise"]
C --> D["2025-03 ~ 06<br/>MCP 时代到来<br/>客户端选 ChatWise/Raycast<br/>Obsidian + 自家 MCP"]
D --> E["2025-07 ~ 12<br/>Claude Code 主战场<br/>Skills + Sub-agents + MCP<br/>本地模型边缘化"]
E --> F["2026-01 ~<br/>OpenCode/newtype-os<br/>多 Agent 编排<br/>To Agent 接口"]
Each era has different tool combinations.Lao Huang's key insight: Tools will be replaced by new tools, but the logic for connecting these tools (API Key / local port / MCP) always remains the same.
1. Lao Huang's current (2025-04~) commonly used combo
From post 1176:
"Gemini + Grok + Raycast + Cursor + NotebookLM, this combo is extremely powerful."
Later additions (posts 2148, 2535):
- Claude Code as a personal AI OS framework (4 major stacks: Skills / MCP / Slash commands / Sub-agents)
- newtype OS / newtype-cli as a Multi-Agent content production team
Eliminated tools (recommended in 2024 but abandoned by Lao Huang after 2025):
- AnythingLLM (downgraded from core to "backup + trial") — Agent features are elementary, RAG performance is mediocre
- Open WebUI (recommended early on, no longer mentioned later)
- LobeChat (used for early deployment of newtype.pro, no longer used later)
- ChatWise (frequently recommended in the first half of 2025; a GitHub issue in 2026-02 showed it stopped being updated, which Lao Huang tacitly accepted)
Retained tools (used continuously since 2024):
- Ollama(local model runtime, the stable king)
- LM Studio(leveraging MLX performance advantages on Mac)
- Obsidian(notes + AI plugins + self-built MCP)
- Cursor(2024-09 → present, main tool for Vibe Coding)
2. Local model runtime layer: Ollama vs LM Studio (essence from post 53, 2024-04)
Lao Huang's classic "misc sorting" post defined the standard for 2024-2025:
| Tool | Intuitive difference | Advantages |
|---|---|---|
| Ollama | Command line | Many models (can download any GGUF) / automatically releases VRAM after 5 minutes by default |
| LM Studio | Graphical interface | High integration (search + download + load + chat) / runs multiple models simultaneously / supports MLX |
Core distinction:
- Client mode: Runs large models on its own + comes with a built-in frontend chat interface
- Server mode: Only runs models, opens a port for other applications (Anything LLM / Open WebUI / Python scripts) to call
Lao Huang's practical choice:
| Scenario | Which to choose |
|---|---|
| Quick start for beginners | LM Studio (graphical interface, one-stop from model search to chat) |
| Mac M-series chips + pursuing performance | LM Studio (MLX optimization) |
| Running Multi-Agent frameworks (AutoGen/CrewAI) | LM Studio (supports running multiple models at the same time, each Agent configured with a different model) |
| Server backend | Ollama (standard port 11434, best ecosystem compatibility) |
| Downloading models directly from Hugging Face | Ollama (supported since 2024-10, post 544) |
| Long-term running | LM Studio (does not release VRAM when open; Ollama releases after 5 minutes by default, requires setting OLLAMA_KEEP_ALIVE=-1) |
Important environment variables(Ollama):
OLLAMA_HOST=0.0.0.0: Listen globally (default only listens on 127.0.0.1, needs to be changed for remote access)OLLAMA_KEEP_ALIVE=-1: Do not automatically release VRAM (resident)OLLAMA_MODELS=<path>: Change model storage location (default is C drive)
When to use:
- Beginners → just install LM Studio, that's enough
- Running Multi-Agent / Python scripts → must use Ollama or LM Studio server mode
- Mac M-series chips → prioritize LM Studio + MLX models
3. Frontend chat layer (2024-2025 evolution)
Early 2024 (2024-03~05): AnythingLLM dominated
- Integrated all backends such as Ollama / LM Studio / OpenAI
- Supported RAG (default vector database LanceDB)
- Multiple Workspaces (each workspace with different models + different documents)
- Query Mode (answers only based on documents, no free generation)
- Launched Agent feature in 2024-04 (post 111), but it was elementary
Mid-2024 (2024-06~08): Open WebUI became popular
- Docker deployment / gave Ollama a ChatGPT-style interface
- Built-in RAG + account management
- Problems: Accounts created by admin default to pending status (post 21); search enhancement using searXng easily returns 403 errors
Late 2024 (2024-09~12): Brave + built-in AI
- Select text in browser → right-click to invoke AI (Ollama)
- Lao Huang said he "quite likes it" and hoped custom options would be added
First half of 2025: ChatWise / Cherry Studio (made by Chinese developers)
- Post 816 (2025-02): "Both are made by Chinese developers. Both provide API and local access. I personally prefer ChatWise, it's much cleaner"
- Post 817: The new version of ChatWise supports Jina DeepResearch
- Post 1176 (2025-04): ChatWise only has 2 uses — backup + trial
Late 2025: Desktop clients gradually gave way to Claude Code / Codex
- Post 1586 (2025-08): "Using Claude Code shows the possibility of relatively general-purpose Agents"
- Post 2032 (2026-01): "newtype Profile integrated into OpenCode/Claude Code"
When to use:
- Multi-model comparison → Cherry Studio or ChatWise (lightweight switching + backup)
- Multi-Agent Workflow → go directly with Claude Code + Sub-agents (late 2025 direction)
- Knowledge base aggregation → AnythingLLM (local) or NotebookLM (cloud)
4. Knowledge base layer: from tool stacking to NotebookLM (2024-2026)
2024 tool free-for-all
| Tool | Evaluation |
|---|---|
| AnythingLLM | General-purpose aggregator, easy to get started, RAG performance mediocre |
| Open WebUI + Ollama | Built-in RAG, suitable for fully local deployment scenarios |
| MaxKB | Suitable for enterprise internal training (post 149), one-click installation via iPanel |
| RAGFlow | Diverse document parsing, better performance than AnythingLLM (comparison in post 490) |
| QAnything | NetEase Youdao, cloud version works well, local deployment is painful |
| GraphRAG | Microsoft, global Q&A, but extremely high cost (post 360: $11 per single document single Q&A) |
2025 simplification: ultimate personal knowledge base solution (essence from post 2201)
Lao Huang's two-sentence summary given in 2026-01:
Your own stuff → Obsidian (local-first) Other people's stuff → NotebookLM (performance first)
Reasons:
- Don't trust any cloud for your own notes; every note in Obsidian is a Markdown file,extremely AI-friendly
- Other people's materials (papers/reports/video transcripts) go to NotebookLM,Source-Grounding is strict and it doesn't make things up
See Full Evolution of newtype Knowledge System and Life OS(to be written, next wiki).
Key points for RAG performance optimization (accumulated from multiple answers)
| Optimization point | How to do it |
|---|---|
| Document preprocessing | Convert PDF to Markdown (use Marker / MinerU / Markitdown) / remove vague expressions / make each section self-contained |
| Embedding model | For Chinese scenarios:nomic-embed-text or mxbai-embed-large; for enterprise use OpenAI text-embedding-3-large |
| Chunk size + overlap | Too small → cuts off semantics; default 1000-token chunk is generally OK |
| Citations check | Supported by AnythingLLM etc., check if the cited text chunks are correct |
| Model selection | RAG has non-trivial reasoning requirements for models; 7B is not enough, prioritize 14B+/Qwen Max/Claude |
5. Obsidian ecosystem (Lao Huang's favorite)
Obsidian core plugins
| Plugin | What it does |
|---|---|
| Copilot(Lao Huang's main recommendation) | AI chat + Vault QA (RAG based on notes) |
| Text Generator | Text generation, can connect to local large models |
| Web Clipper + Interpreter | Automatic clipping + AI summarization (supports Ollama) |
| Smart Connections | Semantic connections between notes |
Copilot setup key points (essence from post 218)
- Default Model: Choose OpenRouter / OpenAI / Ollama
- Embedding:
text-embedding-3-large(OpenAI) - Local access: For LM Studio use
localhost:1234/v1; for Ollama configure via local command line - Vault QA: Change mode from Chat → Vault QA, let the plugin embed all notes to build an index
Lao Huang's evolution (post 1596, 2025-08)
- Obsidian is used to store notes(fully localized, Markdown for AI)
- Do not write notes in Obsidian(miscellaneous features are distracting), use Ulysses to write then paste over
- Delete all AI plugins in Obsidian, switch to self-built Obsidian MCP + Claude
- "Note-taking itself is just those things; I think more about how to further purify and refine my own stuff"
See newtype · MCP Protocol and Ecosystem Practice §2.1 Obsidian MCP.
6. Lao Huang's hardware practice (M4 Mac mini / 3060)
Mac mini M4 as a lightweight AI server (post 677, 2024-12)
Configuration: M4 Mac mini + 24G unified memory Model test (Qwen 2.5):
| Precision | Model size | Tokens/second | Evaluation |
|---|---|---|---|
| Q4 | 7B | 20 | Smooth |
| Q4 | 14B | 10 | Acceptable / Lao Huang's final choice |
| Q8 | 7B | A notch slower | Forget it |
| Q8 | 14B | A notch slower | Forget it |
Ollama configuration:
OLLAMA_KEEP_ALIVE=-1(resident)OLLAMA_HOST=0.0.0.0(LAN listening)
iOS App: Enchanted (smooth)
DeepSeek-R1 cross-device performance (post 783, 2025-02)
| Device | Model | Tokens/second |
|---|---|---|
| 3060 (16G) | 7B | 46 (very smooth) |
| 3060 (16G) | 8B | 44 |
| 3060 (16G) | 14B | 26 (acceptable) |
| M4 Mac mini (24G) | 7B | 19 |
| M4 Mac mini (24G) | 8B | 17 |
| M4 Mac mini (24G) | 14B | 10 |
Conclusion: The advantage of Mac mini is low power consumption; for performance use PC + graphics card.
Mac vs Win choice (posts 1015, 1269)
"No need for comparative analysis, just look at the product launches of companies like OpenAI, they all use Mac. "
But for daily local use / running Python / running Ollama,the difference between Mac and Win is not that big. The advantage of M-series chips lies in MLX.
7. Mobile Tools (2024–2025)
| Platform | Tool | Notes |
|---|---|---|
| iOS | Enchanted | Lao Huang’s top recommendation; connects to Ollama smoothly |
| iOS | Apollo | $5, supports custom backends (local / OpenRouter / self-hosted services) |
| iOS | Typingmind(web shortcut) | Supports multiple models |
| Universal on mobile | Official apps (ChatGPT / Claude / Gemini) | The simplest option / Lao Huang’s recommended starting point |
| Mac terminal | Ghostty | Used at OpenAI’s launch event; Lao Huang knows it but usually uses the Cursor terminal |
Lao Huang’s practice(Post 777):
- Do not chase a “full mobile solution”
- On iOS, use Enchanted to connect to Ollama on the Mac mini at home
- For urgent use outside → ChatGPT official app / Gemini official app
8. Key Tool Combination Formulas
Fully local configuration (data never leaves the device)
Ollama(后端) + AnythingLLM/Open WebUI(前端) + Llama3/Qwen2.5(模型)
+ nomic-embed-text(嵌入) + LanceDB(默认向量库)
Applicable: sensitive data such as legal documents/contracts/financial reports / offline scenarios
Personal hybrid configuration (Lao Huang, mid-2024–2025)
Obsidian(笔记) + Copilot 插件(本地+云端切换)
+ OpenRouter(模型聚合) + Cursor(代码)
+ Perplexity(搜索)+ Anything LLM(资料库)
Lao Huang’s current setup (2025-04+)
Gemini + Grok + Raycast + Cursor + NotebookLM
+ Claude Code(配 Skills/MCP/Sub-agents)
+ Obsidian + 自家 Obsidian MCP
+ Prompt House + Prompt House MCP
Mac mini AI server configuration
Mac mini M4 24G + Ollama(Qwen 2.5 14B Q4)
+ OLLAMA_KEEP_ALIVE=-1 + OLLAMA_HOST=0.0.0.0
+ iOS Enchanted 连接局域网
9. Common Configuration Pitfalls (categorized)
Ollama series
- Docker installation error → Check M-chip compatibility (Post 127)
- Model downloads filled up the C drive → Environment variable
OLLAMA_MODELS=<path>(Post 74 featured) - Chinese models produce nonsense → Modify modelfile.txt to add TEMPLATE and PARAMETER (Post 42)
- GPU not working → Check GGML_CUDA / Layer offload (adjustable in LM Studio settings)
- GPT4ALL? → Not recommended by Lao Huang: too few models + crude product
AnythingLLM series
- Document import error: fetch failed → Switch Embedding (the default Hugging Face one is unresponsive) (Post 30)
- Cannot change Workspace file location → The current version may not support it
- Answers do not reference the original text → 1) Model not strong enough (test with GPT-4) 2) Chunk size/overlap too small 3) Embedding model does not support the language
Open WebUI series
- Newly created accounts default to pending → Admin changes Default User Role to User (Post 21)
- Cannot connect after restart → Ollama did not start
LM Studio series
- How to use local models → Start server mode, port 1234 (default)
- How to use MLX models → Check the MLX version when searching
- Phone cannot connect → The software may not support phone connections (Ollama + Enchanted is a compatible path)
10. Tool Selection Decision Tree
你的核心需求是什么?
├── 单纯聊天 / 多模型对比
│ └── ChatWise / Cherry Studio / Raycast(Mac)
├── 跑本地大模型
│ ├── 新手图形化 → LM Studio
│ ├── 服务器/给其他 App 调用 → Ollama
│ └── Mac M 芯片追极致 → LM Studio + MLX
├── 知识库
│ ├── 自己笔记 → Obsidian + Copilot / Obsidian + MCP
│ ├── 别人资料(论文/报告)→ NotebookLM
│ ├── 团队/企业 → MaxKB / RAGFlow
│ └── 完全本地+敏感数据 → Ollama + AnythingLLM
├── 编程
│ ├── 主力 → Cursor / Claude Code
│ └── 长上下文 → Augment Code(老黄没用过,听过)
├── 行业 / 市场研究
│ └── Gemini Deep Research / Perplexity / Grok
├── 多 Agent 编排
│ └── Claude Code + Sub-agents / OpenCode + newtype-os(老黄做的)
└── Workflow 自动化
└── n8n(配 MCP Server Trigger 转 MCP)/ Dify
11. Things Recommended in 2024 That No One Mentions Now (obsolete)
| Tool | 2024 recommendation time | Current status |
|---|---|---|
| LobeChat | 2024-03 giveaway | Lao Huang stopped using newtype.pro long ago |
| Open WebUI | 2024-03 | Succeeded by ChatWise / Cherry Studio / official apps |
| Phidata | 2024-04 | Succeeded by CrewAI / Claude Code |
| LLocalSearch | 2024-04 | Succeeded by NotebookLM / Perplexica / Perplexity |
| GraphRAG | 2024-07 | Too costly ($11 per Q&A), succeeded by Source-Grounded NotebookLM |
| LightRAG | 2025-02 | Succeeded by ordinary RAG + high-quality Embedding |
| Perplexica | 2024-08 | Succeeded by Perplexity / NotebookLM |
| MaxKB / RAGFlow | 2024 ~ 2025-02 | Retained only for enterprise scenarios |
| AnythingLLM | 2024-2025-04 | Downgraded to “backup + early access,” not a primary tool |
| ChatWise | 2025-02-2026-02 | GitHub issues indicate development has stopped; future unclear |
Why this obsolescence pace is worth watching:
- Models get stronger → application-layer “middleware” gets eaten
- MCP emerges → client-integrated RAG/Search is replaced by external MCP plugins
- Agent systems like Claude Code / Codex emerge → Workflow tools (Dify/Coze) are marginalized
12. Lao Huang’s Principles for Choosing Tools (summary)
| Principle | Source |
|---|---|
| Stick to the best; domestic Chinese AI is not worth considering | Post 2063 (“Since you’re going to learn and use it, just focus on the best”) |
| Mac is standard at OpenAI events, but Windows runs AI fine too | Post 1015 |
| In the AI era, API calls > local deployment(unless data is sensitive) | Posts 593, 874 |
| Local large models’ coding ability can’t catch up to Claude 3.5 Sonnet | Post 593 |
| Embedding models matter; get OpenAI embeddings working first, then downgrade to local | Post 886 |
| AnythingLLM is easy to use, but its Agent is still rudimentary; use Claude Code for production | Posts 449, 1586 |
| Pick the most popular tools; if unsure, don’t tinker | Post 70 |
| For an AI knowledge base, the most important thing isn’t AI—it’s knowledge | Post 717 |
13. Anti-patterns
| ❌ | ✅ |
|---|---|
| Hearing that tool X is good, installing it, and recommending it without trying | Judge after using it yourself,and compare it with similar tools you’ve used |
| Blindly pursuing “fully local” | In most scenarios APIs are cheaper and work better; go local only when data is sensitive |
| Running a 70B local large model on a Mac mini | 24GB VRAM isn’t enough; it produces weird results—just use the cloud |
| Stuffing all materials into the knowledge base | Only put in clear, self-contained, high-quality content in QA form |
| Using a 7B model for RAG and expecting GPT-4-level results | The model’s reasoning power is insufficient; validate with GPT-4 first, then downgrade to local |
| Using one AI tool for everything | Divide work among v0 / Claude / Cursor (see newtype · Vibe Coding: Building Independent Products in Practice §4) |
| Treating “installing tools” as the goal | Installing tools is a means;forming a reusable SOP is the goal |
| Obsessing over a discontinued tool | Tools get replaced;the core is the integration logic (API + ports + MCP) |
References and Further Reading
Main references: All 178 posts in the tools-stack topic bucket (2024-03 to 2026-05), no sampling. Covered post numbers include (complete full list): 11, 12, 15, 20, 21, 28, 29, 30, 32, 33, 34, 36, 39, 40, 42, 43, 44, 47, 52, 53, 56, 57, 64, 67, 68, 70, 73, 74, 80, 82, 83, 87, 90, 93, 95, 96, 98, 99, 101, 110, 111, 116, 122, 125, 127, 131, 133, 136, 140, 142, 147, 149, 155, 160, 161, 163, 164, 166, 172, 179, 180, 196, 214, 218, 220, 242, 259, 269, 276, 284, 305, 315, 353, 356, 360, 363, 377, 401, 422, 423, 431, 437, 444, 445, 449, 460, 469, 476, 490, 510, 526, 535, 544, 552, 563, 570, 573, 593, 612, 623, 656, 668, 673, 677, 690, 717, 766, 767, 770, 771, 772, 777, 780, 782, 783, 786, 791, 792, 793, 794, 795, 796, 816, 817, 824, 830, 831, 836, 841, 874, 882, 883, 921, etc.
Companion wiki:
- newtype · MCP Protocol and Ecosystem in Practice — modern tool combinations are connected via MCP
- newtype · Prompt / Context Engineering in Practice — tools are the carrier; Prompt is the content
- newtype · Vibe Coding: Building Independent Products in Practice — how to use Cursor / Claude Code
- (TODO)newtype Knowledge System and the Full Evolution of Life OS — Obsidian + NotebookLM end-state solution
- (TODO)newtype · AI Coding in Practice and Tool Stack Evolution — Cursor / Claude Code / Cline / v0 / Codex
History
- 2026-05-16 v1.0 created — Phase 1.1 wiki #6. Based on all 178 posts in the tools-stack topic bucket. No sampling. Organized across 13 dimensions: evolution timeline / Lao Huang’s current combination / Ollama vs LM Studio / evolution of the frontend chat layer / knowledge base layer / Obsidian ecosystem / hardware practice / mobile / combination formulas / configuration pitfalls / decision tree / obsolescence list / selection principles / anti-patterns.