Everyday AI tools
It is hard to write code without AI these days. I currently use this stack:
| Scenario | Tool | How I use it |
|---|---|---|
| IDE | Cursor | Write and edit files, and inspect pages daily |
| CLI | Claude Code | Long, well-bounded tasks that can run asynchronously |
| App | Codex | Desktop app; an old account with a Turkish gift card costs a little over ¥70 per month |
| Intelligence | Codex Radar | Track quotas, model quality, degradation, and community sentiment |
| Switching | CC Switch | Change providers for Claude Code |
| API | DeepSeek | Cheap, directly reachable from China, and a fallback for CLIs and scripts |
For an earlier guide to Cursor, see Cursor AI User Guide.
Cursor
Cursor is still the most comfortable choice for daily work. It is based on VS Code, and multi-file Agent edits, completion, and chat all live in one window. I use Agent mode most: explain the requirement clearly, let it edit, then review the diff myself. Completion can sometimes take over your train of thought, so enable it when needed. Put project conventions in .cursor/rules; it works better than repeating them in chat.
The official Pro plan is not cheap. In China, common options include buying accounts on Xianyu or sharing a plan. I have done this too, mainly to get the IDE running first. Check reviews and transaction history, and do not tie your only work environment to an unknown shared account. Keep it separate from your primary account. You bear the risk of bans and expiration. The official subscription is safer if you can accept the price.
Claude Code
Anthropic’s command-line Agent, run as claude in a project directory. It is good for handing off a complete task: fix a bug, add tests, or update a page in the existing style, while you go do something else.
npm install -g @anthropic-ai/claude-codeOfficial documentation: Claude Code.
The split with Cursor is simple: use Cursor when you are still in the editor and need to watch the UI; hand a well-bounded asynchronous task to Claude Code. It reads CLAUDE.md and AGENTS.md in the repository, so conventions are more useful in files than in chat history.
Codex
I use it as an app: download, sign in, and open repositories from the desktop application. Repository: openai/codex
The subscription uses Apple gift cards from the Turkish region: top up first, then subscribe. My tier costs a little over ¥70 per month, roughly half the official price. New accounts no longer get this price; it is a renewal on an old account, so do not use it as the expected price for a new account. Use a reliable gift-card channel and do not mix the region or Apple ID with your main Chinese-region account.
I use it as a backup and second perspective alongside Claude Code: run the same task in both, and switch when Claude is unreliable or I want an OpenAI-style model. Because the app subscription goes through Apple, there is no need to configure a CLI provider for Codex.
Codex Radar
For quota status, model quality, and community sentiment, I check Codex Radar. It also has ready-made notes on enabling Max reasoning and connecting DeepSeek to Codex.
CC Switch
Before CC Switch, changing Claude Code providers meant editing ~/.claude and environment variables. It quickly became messy with multiple channels. CC Switch is a desktop app where providers, MCP, and Skills can be switched with a click. GitHub: farion1231/cc-switch
I add the official provider, DeepSeek, and GLM, then enable them from the tray. Claude Code can usually switch without restarting. The point is not a smarter model; it is switching the backend under the same CLI: use DeepSeek when quotas are tight, GLM when I want a domestic provider, then switch back to the official provider.
DeepSeek developer platform
deepseek.com is the chat product, while platform.deepseek.com is the API. Use the latter as the backend for tools, scripts, and Claude Code.
The APIs are compatible with OpenAI (https://api.deepseek.com) and Anthropic (https://api.deepseek.com/anthropic); change base_url and the key to connect. Common models include deepseek-v4-flash and deepseek-v4-pro. It is cheap and directly reachable from China, making it a fallback when official quotas are tight. Store the key in an environment variable. Documentation: DeepSeek API Docs
How I combine them
- Open the repository and start with Cursor.
- When the task gets longer and can run asynchronously, hand it to Claude Code.
- Use CC Switch to change channels, with DeepSeek as the inexpensive fallback.
- For a different perspective or when Claude is not working well, open the Codex App; check Codex Radar for quota and model-quality reports.
- For a resident local agent with WeChat as an entry point, use Hermes Agent.
- To expose a CLI subscription as an API for other clients, use CLIProxyAPI.
Tools will change. The stable habits are to put conventions in rule files, review your own diffs, and never confuse “it runs” with “it is correct.”