Assistant: Support custom URLs for providers such as OpenAI

[juliasilge]

Originally posted by @cpilat97 in #7985:

Hello there! So far have had pretty good experience with the use of copilot in-line (that's what my company uses and we have accounts to). For use of the claude models, we have a custom url for openai that we use. So when i'm using ellmer::chat_openai() i supply a base_url and my api_key.

I have something similar with chattr in my .Rprofile to point to those.

I'm not sure of the level of support to do so, but as y'all explore additional models beyond anthropic, I just wanted to throw this out there as potential inclusion to point to specific urls. I don't know if how the chats are currently implemented would support this but would be a great additional feature.

Thanks for the continued work you put into this!

[chumpalump]

We'd love to see support for custom URLs for OpenAI (and others as time goes on). At my company we need to access openai through a custom endpoint. It's a security layer, but it acts like a passthrough proxy. That means exposing the api endpoint to the customizable settings would be the only needed change.

[kv9898]

This also helps open doors to OpenAI-compatible services such as DeepSeek. Really looking forward!

[jackojacko05]

👍 +1 This is exactly our situation too.

Our company host LiteLLM in Anthropic-compatible mode behind https://llm.internal.example.com/v1.
Positron 2025.07.0-204 cannot pick up ANTHROPIC_BASE_URL, so the Assistant always calls https://api.anthropic.com and returns 401.

Being able to set a custom base URL (and header name) is essential for enterprise environments where direct Anthropic / OpenAI access is blocked. Happy to test any PR or Nightly build!

[carldelfin]

Recently used this feature in Zed and was sad to see it wasn't possible to do in Positron. Like others above me, my company makes use of a custom URL for OpenAI - hopefully this is not too far into the future!

[ZhimingYe]

I hope the Positron team will give further consideration to the priority of this issue. Although I understand the team wants the AI to provide a first-class experience (and it is indeed currently optimized only for Anthropic), opening up some customization options—so long as this doesn’t add to the team’s workload—would help achieve an open and free IDE experience (not tied to any third-party company). Anthropic can indeed be too expensive at times and has regional restrictions (for me, in most cases it can only be accessed via OpenRouter). Now Deepseek also supports the Anthropic API (https://api-docs.deepseek.com/guides/anthropic_api), and I believe this is the future trend — API convergence, even if different models are offered. Thank you!

[kv9898]

Guys, deepseek seems to work fine so long as we change the API endpoint. Even the agent mode works. I created PR #9241 to add support for Anthropic and OpenAI-compatible models.

If you would like to try it out, I have made a Windows x64 build of the latest Positron with this feature.

User Instruction

To enable this feature, add to settings.json:

"positron.assistant.enabledProviders": ["anthropic-compatible", "openai"]

And then you can add Anthropic and OpenAI-compatible models through the normal UI:

[kv9898]

@juliasilge One thing I noted is that custom urls for OpenAI models almost work out of the box once OpenAI models are activated in settings.json and added to the package.json of positron-assistant. Is there a particular reason that this 'existing' support gets locked?