Open AI API compatible ?

[Greatz08]

I used find and search on main page and didnt found any openai reference so thought to ask.

[ryanjsalva]

At this time, we're optimizing Gemini CLI for Gemini models, and not building direct support for other LLM providers.

[niutech]

[gpetrov]

Seems somebody created an openAI compatible proxy server for Gemini CLI, still needs a lot of improvements though:

https://huggingface.co/engineofperplexity/gemini-openai-proxy/tree/main

[niutech]

@gpetrov This is the other way round - it provides OpenAI-compatible API for Gemini 2.5 Pro - whereas we need a Gemini-CLI-compatible API for OpenAI-compatible local LLM server. In other words:
gemini-cli -> http://localhost:1234/v1internal:generateContent (proxy) -> http://localhost:11434/v1/chat/completions (ollama)

[marty-sullivan]

@Greatz08 The latest release of LiteLLM supports gemini cli. So you can configure a local proxy, configure any model providers you want, and then use them in Gemini CLI. Works great

https://docs.litellm.ai/docs/tutorials/litellm_gemini_cli

[Greatz08]

@marty-sullivan thanks for the info :-)). Will test soon

[retanoj]

litellm_gemini_cli meets some compatibility problems. Which like, Tool "undefined_tool_name" not found in registry.

[marty-sullivan]

@retanoj if you submit an issue to their github, they usually fix issues like this quickly

[acoliver]

We also forked Gemini-cli and added multi-provider for openai and Anthropic https://github.com/acoliver/llxprt-code

We keep up with gemini-cli released and features.

It works great with GPToss and Qwen3-coder. The proxy solution sounds good in theory, but the Gemini Code Assist team is right that providing a good experience for different models is a lot of work. So, the readManyFiles tool, if a model with a less than 1m context window requests 500 files, what happens? In gemini-cli it dies and potentially blows the heap. In llxprt-code it gets a warning not to do that, and the response is limited.

I think the Gemini-cli team made the right choice: focusing on the best possible Gemini 2.5 Pro experience upstream while community-driven projects downstream.

[laiso]

I’m using Qwen Code (Gemini CLI Fork) and OpenCode and Codex CLI for this purpose.

I believe it’s best for Gemini CLI to stay optimized for Gemini models. What I’d really like to see is an agent that makes the most of Gemini’s unique model characteristics and shows off its full potential.

[niutech]

How is Gemini CLI optimized for Gemini models? Aren't all LLMs prompted in standard English?

[laiso]

That’s true, the current dependency isn’t that strong. As for the tools, web_search and web_fetch won’t work if you switch from the Gemini API to another model. But since replacement implementations are possible, that’s only a small gap. What I’m hoping for is the kind of optimization you see in Codex’s responses API or Claude Code’s beta flags.

[acoliver]

Llxprt-code has the web-tools and web-fetch working. We isolate them and use your gemini key or oauth for them. Your chosen model for others. We also add claude style todo list tools, atomic function calls and more. Additionally we tighten the tool calls giving them limit options for both items and tokens. Without that youre limited to reasoning models with large context Windows and get poorer results.

[ttttmr]

I wrote a proxy service that allows Gemini CLI to directly use OpenAI-compatible APIs.
https://github.com/ttttmr/openproxy