| name | local-llm |
| description | Use OpenAI-compatible endpoints (Ollama, vLLM, LM Studio) with this studio: base URL, model name, limits, and anti-hallucination rules for wx/tt APIs. |
| argument-hint | [no args] |
| user-invocable | true |
| allowed-tools | Read, Glob, Grep |
Configure the client
Point any OpenAI-compatible client at your local server:
- Ollama:
baseURL http://127.0.0.1:11434/v1，apiKey 通常为占位字符串（依客户端要求）。
- vLLM / LM Studio：使用其文档中的
/v1 地址与端口。
Always set:
- A model string exactly as the server exposes it.
- Reasonable max tokens / timeouts for long Skills; split work when output truncates.
Prompting norms (this repo)
- Load
STUDIO.md at session start when acting as a coding agent.
- For WeChat / Douyin work, require citations to official docs or in-repo typings before accepting new API usage.
- Treat
.claude/skills/*/SKILL.md as workflow SOPs even when the UI has no Slash menu: the skill name field is the stable command id.
Cursor / IDE notes
Set the editor’s custom OpenAI-compatible provider to the same baseURL. Cursor model dropdown labels vary by version; the important part is the HTTP compatibility, not the product name.
