Thinchat — AI skill for Claude Code

thinchat

[](https://github.com/seokhoonj/thinchat/actions/workflows/check.yml) [](https://pypi.org/project/thinchat/) [](https://pypi.org/project/thinchat/) [](https://github.com/seokhoonj/thinchat/blob/main/LICENSE)

**English** | [한국어](README.ko.md)

One interface for four LLM providers — **claude, openai, gemini, ollama** — that treats **not leaking your API key** as part of the job.

Name a provider, then call it: completion (whole, streamed, or JSON-structured) and, where the provider has one, embeddings — each with an async twin, over the openai and anthropic SDKs. No gateway, no router, no cost tracking — just the calls, plus the parts you'd otherwise hand-roll:

• Leak-safe keys. Your key is a credbox Secret (masked in logs and tracebacks), revealed only at the SDK call. On a failure, thinchat severs the SDK error's chain and scrubs the message so the key never rides out on a traceback — and it pins each provider's official endpoint, so an environment variable can't redirect a resolved key to another host.
• One shared credential store. thinchat set claude saves a key once to a 0600 store that every session — and every sibling tool — resolves from; or just use _API_KEY from the environment (the library core reads no file unless you ask).
• The fiddly bits, handled. Structured output parses tolerant JSON into a dict; embeddings return one alignment-checked vector per input; RateLimitError carries retry_after; supports() reports a provider's capabilities before you call.

1. Install

pip install thinchat

Requires Python 3.11+. Both provider SDKs (openai and anthropic) come with it, so every provider works out of the box; each is imported lazily the first time y