Add support for oMLX and generic OpenAI-compatible API as AI providers

What is the improvement or update you wish to see?

Currently, GitButler supports AI providers including OpenAI, Anthropic, Ollama, and LM Studio for features like AI-generated commit messages and branch names. However, macOS users on Apple Silicon who use oMLX — a native MLX-based local LLM inference server — have no way to integrate it as an AI provider.

There are two proposed improvements:

1. Add native oMLX support as a dedicated AI provider option, similar to how Ollama and LM Studio are supported. oMLX is an open-source, Apple Silicon-optimized LLM inference server that exposes an OpenAI-compatible API endpoint, leverages Apple's MLX framework for high-performance inference, and features a native macOS menubar app for easy management.

2. Add a generic "Custom OpenAI-compatible API" provider option, allowing users to specify any base URL and API key for any inference server that implements the OpenAI Chat Completions API specification. This would future-proof the AI provider selection and avoid the need for new provider-specific code every time a new local inference server gains popularity.

Is there any context that might help us understand?

• oMLX (https://github.com/jundot/omlx) is a local LLM inference server specifically built for Apple Silicon (M1/M2/M3/M4+), using Apple's MLX framework. It exposes an OpenAI-compatible REST API, making integration straightforward. It supports continuous batching and two-tier KV caching (memory + SSD), and comes as a signed/notarized DMG — no CLI setup required.

• Many emerging local inference tools (e.g., oMLX, llama.cpp server, vLLM, LiteLLM, Jan, etc.) already implement the OpenAI-compatible API spec. Adding a generic "Custom Endpoint" option with configurable base URL, model name, and API key would unify support for all these tools without requiring individual integrations.

• This is particularly valuable for macOS Apple Silicon users where oMLX significantly outperforms alternatives like Ollama for MLX-native model inference due to deeper hardware optimization.

• Reference: The existing Ollama integration (PR #3741) demonstrates that custom endpoint support is architecturally feasible within the current AI client interface.

Does the docs page already exist? If so, please link to it.

https://docs.gitbutler.com/features/branch-management/ai-assistance