Supporting Vertex-based models using google auth

[etrippler]

Eg, something like:

{
  "$schema": "https://opencode.ai/config.json",
  "theme": "openai",
  "provider": {
    "vertexai": {
      "npm": "@ai-sdk/google-vertex",
      "options": {
        "project": "my-project-1",
        "location": "us-central1",
        "googleAuthOptions": {
          "keyFile": "creds.json"
        }
      },
      "models": {
        "gemini-2.5-pro": {},
        "gemini-2.0-flash": {},
        "gemini-2.5-flash": {},
        "claude-4.0-sonnet": {},
        "claude-4.0-opus": {}
      }
    }
  }
}

And actually this isn't fully accurate, @ai-sdk/google-vertex works for Gemini models but @ai-sdk/google-vertex/anthropic is needed for Claude models hosted on Vertex.

Alternatively, what I did for now is simply use my litellm proxy gateway running in a seperate process with the following config:

{
  "$schema": "https://opencode.ai/config.json",
  "theme": "openai",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://localhost:4000/v1",
        "apiKey": "sk-demo-1234"
      },
      "models": {
        "gemini-2.5-pro": {},
        "gemini-2.0-flash": {},
        "gemini-2.5-flash": {},
        "claude-4.0-sonnet": {},
        "claude-4.0-opus": {},
        "claude-3.7-sonnet": {},
        "o4-mini-high": {},
        "o4-mini": {},
        "o3": {}
      }
    }
  }
}

The only problem with this is you don't get token/cost tracking data. I think a fair number of people do this, so it could be nice to simply flag somehow that the model you're passing through is the "official" model from models.dev?

[NachoVazquez]

Oh yes, I need this, all my credits are in Vertex.

@thdxr, happy to help with the implementation. I won't jump into it right now, as you're assigned, but please let me know.

[admSla99]

I would also like to see Vertex AI support for Anthropic models. I believe most large companies already provide their employees access to LLMs through Vertex AI due to corporate policies. What is the current status of this feature? I look forward to helping with this.

[fredrikaverpil]

This worked for me, targeting a GCP project in Vertex AI (aiplatform.googleapis.com):

{
  "$schema": "https://opencode.ai/config.json",
  "theme": "system",
  "autoupdate": true,
  "provider": {
    "google-vertex": {
      "models": {
        "gemini-2.5-flash": {},
        "gemini-2.5-pro": {}
      },
      "options": {
        "project": "{env:GOOGLE_CLOUD_PROJECT}",
        "location": "{env:GOOGLE_CLOUD_LOCATION}"
      }
    }
  }
}

Authentication seems to be made under the hood by gcloud auth - or was it the fact that I ran opencode auth and gave it the access token from gclound auth application-default print-access-token? 🤷
I can see from cloud logs I'm making authenticated requests, so I can see that it really works.

[NachoVazquez]

This is interesting, @fredrikaverpil. How would you define the model if you are using Claude through Vertex?

[fredrikaverpil]

@NachoVazquez not sure what you mean. I'm using gemini-cli and various Neovim plugins (codecompanion, avante) as well as opencode with the Gemini models in Vertex, namely flash 2.5 and pro 2.5. The models are specified in the snippet above.

[NachoVazquez]

I'm trying to use Claude Code through Vertex. That's how I use it with Claude Code. I tried following your config, but it doesn't seem to work for me.

...
  "model": "vertex/claude-sonnet-4",
  "provider": {
    "vertex": {
      "models": {
        "claude-sonnet-4": {}
      },
      "options": {
        "project": "{env:GOOGLE_CLOUD_PROJECT}",
        "location": "{env:GOOGLE_CLOUD_LOCATION}"
      }
    }
  },

[NachoVazquez]

Some development is needed, it seems, as stated in the issue description.

[fredrikaverpil]

Claude is not listed under Vertex in https://models.dev, which is what opencode uses.

[NachoVazquez]

Right, there are two PRs related to this, but it would probably not be enough since @ai-sdk/google-vertex is not enough; @ai-sdk/google-vertex/anthropic is needed here.

I will try to put a PR together this weekend.

sst/models.dev#23
sst/models.dev#60

[king-11]

Both the PRs are merged I think we can work on bringing the anthropic SDK into opencode?

[jacobmikesell]

Opened #2028 and confirmed its enough, theres probably a more elegant way, but it gets me off to the races for now!

[cafreeman]

Any updates on this PR? I think this is the only thing between me and being able to use opencode at work 🤞

[jacobmikesell]

@thdxr is this something that qualifies for open source contribution (I'm unsure what the line is for "core features") and/or would you be able to take a look?

[rekram1-node]

more to come on this v soon