LOCAL_ENDPOINT variable does not work

[sammcj]

I wanted to try out Opencode with a number of models I've got running on my Ollama and Llama.cpp servers (using the OpenAI compatible API endpoints).

I exported LOCAL_ENDPOINT=https://ollama.my.internal.network/v1 however when Opencode starts it connects to Github Copilot - even if I have it disabled in the config file.

While I suspect this is a bug - I would also recommend allowing users to set their OpenAI compatible API provider, model and model inference settings the .opencode.json config file so you're not relying on the users shell environment (and potentially clashing with other applications).

[ianks]

Ran into same issue. In my case, I am trying to use LOCAL_ENDPOINT as a proxy url.

[sam99dave]

Ctrl + o gives the option to change model config

Im using llama.cpp, for me its connecting to my local endpoint but giving the following error

got exception: {"code":500,"message":"Assistant message must contain either 'content' or 'tool_calls'!","type":"server_error"}

[obukhovaa]

@sammcj currently the local provider only support following api paths to fetch models:

	localModelsPath        = "v1/models"
	lmStudioBetaModelsPath = "api/v0/models"

but ollama expect /api/tags, so it wouldn't work by default. However you said you have OpenAI compatible endpoints, in this case try to use this instead:

LOCAL_ENDPOINT=https://ollama.my.internal.network

For me solution was to use LiteLLM proxy in between, and then allow LiteLLM handle both local ollama and remote compatible API.

In order to work with LiteLLM properly (auth via token + pass through for vertex, for example) a few changes were required to be made, here they are: #243. Im keeping it in sync, but not sure if any maintainers actually managing this repo at the moment, which is a bit annoying :)