Feature Request: Deepseek V4 via Opencode Go API keys

[mrooney]

Deepseek V4 models are popularly available through Opencode Go subs and it would be awesome to be able to plug in the Opencode Go API key and get the same features and image passthrough support with this subscription!

Thanks for any consideration or workarounds :)

[Vizards]

Thanks, this is a very reasonable direction.

As Copilot gradually opens up BYOK model access, I’m not sure we can keep relying on models included in the Copilot subscription for the vision proxy forever. I may explore adding a “one-click setup” path in the extension for Opencode Go, so users can plug in an Opencode Go API key/endpoint and keep the same experience where possible.

One question for you: for your workflow, which would be friendlier: connecting directly to the official DeepSeek API, or using an Opencode Go subscription? And why?

[AzureZeng]

Thanks, this is a very reasonable direction.

As Copilot gradually opens up BYOK model access, I’m not sure we can keep relying on models included in the Copilot subscription for the vision proxy forever. I may explore adding a “one-click setup” path in the extension for Opencode Go, so users can plug in an Opencode Go API key/endpoint and keep the same experience where possible.

One question for you: for your workflow, which would be friendlier: connecting directly to the official DeepSeek API, or using an Opencode Go subscription? And why?

Maybe it is cheap enough. If you purchased OpenCode Go (10$ per month) and use DeepSeek V4 Flash in the subscription, you will have 60$ credits and you can use without any balance worries.

I also subscribed OpenCode and I only used 10% of the monthly balance, and it is in my heavy usage.

[mrooney]

Which would be friendlier: connecting directly to the official DeepSeek API, or using an Opencode Go subscription? And why?

As AzureZeng says, you get $60 in credits at a fixed monthly cost; it's a great deal and easy to understand. So I'd love to use the API key I already have. It wouldn't really make sense to buy separate Deepseek tokens when I already am not hitting my cap on the subscription I have.

Regarding vision, OpenCode Go also provides access to several model families such as Kimi and Qwen. It looks like Qwen3.5 Plus would be the cheapest for vision. So if you added support for Opencode Go, it would include its own vision fallback. Pretty handy!

I may explore adding a “one-click setup” path in the extension for Opencode Go

That would be amazing!

[Vizards]

@mrooney @AzureZeng Thanks again for the context here.

I decided to first support a generic BYOK vision proxy instead of adding an Opencode Go-specific path immediately. This is now available in v0.6.0: you can run DeepSeek: Configure Vision Proxy from the Command Palette, choose API endpoint, and configure any OpenAI-compatible / Anthropic-compatible vision-capable model with your own key.

Further first-class Opencode Go integration is still something I’m looking into, especially around making setup smoother for Opencode Go users.

[mrooney]

Awesome! So the BYOK is just for the vision proxy currently, is that correct? I assume I cannot use my Opencode Go API key to access Deepseek models in this update?

[Vizards]

Awesome! So the BYOK is just for the vision proxy currently, is that correct? I assume I cannot use my Opencode Go API key to access Deepseek models in this update?

Oh! So you want to enter the Opencode Go baseUrl and apiKey in the extension's settings! Let me try and test the compatibility...

[Vizards]

@mrooney I tested Opencode Go with both deepseek-v4-flash and deepseek-v4-pro, across none / high / max reasoning modes, and the requests were accepted successfully.

Configure main DeepSeek model

Run DeepSeek: Open Settings in VS Code Command Palette, then set the baseUrl to https://opencode.ai/zen/go/v1:

If you prefer configure via json:

"deepseek-copilot.baseUrl": "https://opencode.ai/zen/go/v1"

Then run DeepSeek: Set API Key again and enter your Opencode Go API key.

That's It, You can work with DeepSeek in Copilot, but use the API quota provided by Opencode Go!

Configure a separate vision proxy:

1. Run DeepSeek: Configure Vision Proxy
2. Choose API endpoint
3. Set the endpoint URL to: https://opencode.ai/zen/go/v1/chat/completions
4. Select OpenAI-compatible / Chat Completions（Auto selected）
5. Enter the Opencode Go vision-capable model you want to use. Maybe Qwen3.5 Plus
6. Enter your Opencode Go API key and click Save

So yes, with the current version you should be able to use Opencode Go for both the main DeepSeek V4 model and a separate vision proxy model.

[mrooney]

Wow, it works, you rock! I love the vision test, that's very clever. If you'd like any help / PR to put this in the README or somewhere else, let me know. Eventually, a little setup flow to smooth this out for Opencode Go would be great, but this is super cool for now.

PS - I used qwen3.7-plus for the vision proxy, as it is currently cheaper than 3.6-plus for whatever reason, and 3.5 is not available.