支持配合本地模型么？

[Calmact]

配合本地模型，比如Claude Code对接lmstudio部署的本地1234端口的API服务，是否支持转发？不启动。

[liaohch3]

Short answer: yes, if the local model server exposes the same API shape as the client expects.

For Claude Code, the local endpoint needs to be Anthropic-compatible (for example /v1/messages). You can point claude-tap at the local upstream with --tap-target http://127.0.0.1:<port>; after #105, launched Claude sessions also get ANTHROPIC_BASE_URL injected toward the local tap proxy unless you pass your own --settings.

For OpenAI-compatible local servers such as LM Studio, this probably belongs under Codex/opencode-style client support rather than Claude Code directly. Keeping this open as an enhancement/docs follow-up until we have a tested recipe.

[YoungCan-Wang]

哈哈 我的pr给别人解决问题了

[liaohch3]

@YoungCan-Wang 哈哈，是的，你之前的 PR 已经帮后续用户解决了类似场景的问题。感谢你的贡献，我已经把你加入 README 的贡献者区了。\n\n\n\n再次感谢你对 claude-tap 的改进。

[YoungCan-Wang]

@liaohch3 有需要的话 可以和你一起共建 这个项目我很感兴趣

[liaohch3]

@liaohch3 有需要的话 可以和你一起共建 这个项目我很感兴趣

可以加我微信 liaohch3～

[YanTianlong-01]

Short answer: yes, if the local model server exposes the same API shape as the client expects.简短回答：可以，只要本地模型服务器暴露的 API 格式与客户端期望的一致即可。

For Claude Code, the local endpoint needs to be Anthropic-compatible (for example /v1/messages). You can point claude-tap at the local upstream with --tap-target http://127.0.0.1:<port>; after #105, launched Claude sessions also get ANTHROPIC_BASE_URL injected toward the local tap proxy unless you pass your own --settings.对于 Claude Code，本地端点需要兼容 Anthropic 格式（例如 /v1/messages ）。你可以通过 --tap-target http://127.0.0.1:<port> 将 claude-tap 指向本地上游；在 #105 之后，启动的 Claude 会话也会自动将 ANTHROPIC_BASE_URL 注入到本地 tap 代理，除非你自行传入 --settings 。

For OpenAI-compatible local servers such as LM Studio, this probably belongs under Codex/opencode-style client support rather than Claude Code directly. Keeping this open as an enhancement/docs follow-up until we have a tested recipe.对于兼容 OpenAI 格式的本地服务器（如 LM Studio），这更适合归入 Codex/opencode 风格客户端的支持范畴，而非直接用于 Claude Code。在获得经过测试的可行方案之前，我们先将此问题作为功能增强/文档跟进事项保持开放。

意思是目前对于本地运行OpenAI格式的模型请求还不支持吗？

[YoungCan-Wang]

This should work when the local model server exposes an API shape that the traced client can actually speak.

The important distinction is the client protocol:

• Claude Code expects an Anthropic-compatible API surface such as /v1/messages.
• Codex/OpenAI-style clients expect an OpenAI-compatible API surface such as /v1/responses or /v1/chat/completions, depending on the client.
• LM Studio usually exposes OpenAI-compatible endpoints, so it is more naturally paired with Codex/OpenAI-style clients than Claude Code unless an Anthropic-compatible adapter is added in front.

I think the next useful step here is documentation with tested local-model recipes, not core proxy code. If someone has a specific local server plus client combination that fails despite matching the expected API shape, please open a narrower bug with the exact command and error.