Last updated 2026-05-17

Vertex AI Model Garden — Anthropic

The vertexanthropic provider talks to Claude models hosted on Google Vertex AI Model Garden. The transport layer is Vertex-shaped (regional :streamRawPredict endpoint, OAuth2 access token from Google ADC), but the request body and the SSE event stream are native Anthropic. There is one Vertex-specific quirk — covered below — and one model-id format change.

Picking the right Anthropic surface:

Import path

import (
    "github.com/elloloop/llmrouter"
    "github.com/elloloop/llmrouter/providers/vertexanthropic"
)

Endpoint shape

Vertex serves Anthropic models from a regional, project-scoped endpoint: 

POST https://<region>-aiplatform.googleapis.com/v1/projects/<project>/locations/<region>/publishers/anthropic/models/<model>:streamRawPredict

The provider composes this URL from WithProject, WithRegion, and the model id on the request. You can override the host with llmrouter.WithBaseURL if you front Vertex behind a custom hostname.

Vertex-specific quirks

  • The Vertex version goes in the body. Anthropic on Vertex requires a top-level "anthropic_version": "vertex-2023-10-16" field in the request body — not in a header. The provider injects this automatically; you do not need to set it on ChatRequest.
  • Model id uses @-versioning. Vertex addresses Claude variants as claude-3-5-sonnet-v2@20241022, not the dash-separated form direct Anthropic uses. Pass the Vertex form verbatim in ChatRequest.Model — the provider interpolates it straight into the URL. (The router can translate vendor-neutral ids like claude-3-5-sonnet to the @-versioned form for you via router.ApplyModelTranslation.)
  • Pure HTTP, no google.golang.org/genai dependency. Unlike providers/vertex (which uses the genai SDK for Gemini), this package is intentionally a pure-HTTP provider — it does not import the GCP SDK. Bring your own token source.

Authentication

Vertex uses Google ADC (Application Default Credentials), surfaced here as an OAuth2 access token sent in the Authorization: Bearer header. Two ways to provide one — pick exactly one:

  • WithTokenSource(src TokenSource) — preferred. Your callback is invoked once per CompletionStream call. Implement caching and refresh inside the callback (the standard pattern wraps an oauth2.TokenSource from golang.org/x/oauth2/google).
  • WithAccessToken(token string) — static. Convenient for CI runs where the token is minted out-of-band and lives long enough for the request to complete. Not appropriate for long-lived servers (tokens expire every hour).

llmrouter.WithAPIKey is rejected at New(). Vertex does not accept API keys for Model Garden endpoints, and accepting one silently would misconfigure auth.

Construction

The canonical pattern wraps a Google ADC token source: 

import (
    "context"


    "github.com/elloloop/llmrouter"
    "github.com/elloloop/llmrouter/providers/vertexanthropic"
    "golang.org/x/oauth2/google"
)


func newProvider(ctx context.Context) (*vertexanthropic.Provider, error) {
    creds, err := google.FindDefaultCredentials(ctx,
        "https://www.googleapis.com/auth/cloud-platform")
    if err != nil {
        return nil, err
    }


    return vertexanthropic.New(
        vertexanthropic.WithProject("my-gcp-project"),
        vertexanthropic.WithRegion("us-east5"),
        vertexanthropic.WithTokenSource(func(ctx context.Context) (string, error) {
            tok, err := creds.TokenSource.Token()
            if err != nil {
                return "", err
            }
            return tok.AccessToken, nil
        }),
    )
}

The token source closes over the cached oauth2.TokenSource, so refreshes happen transparently inside the callback.

Capabilities

Because the body and SSE event stream are native Anthropic, every feature of the Anthropic provider works without code changes — typed tool use, extended thinking deltas, prompt caching, multipart content, mid-stream errors, structured outputs via forced tool-use, tool-result messages.

Full example

package main


import (
    "context"
    "fmt"
    "log"


    "github.com/elloloop/llmrouter"
    "github.com/elloloop/llmrouter/providers/vertexanthropic"
    "golang.org/x/oauth2/google"
)


func main() {
    ctx := context.Background()


    creds, err := google.FindDefaultCredentials(ctx,
        "https://www.googleapis.com/auth/cloud-platform")
    if err != nil {
        log.Fatal(err)
    }


    p, err := vertexanthropic.New(
        vertexanthropic.WithProject("my-gcp-project"),
        vertexanthropic.WithRegion("us-east5"),
        vertexanthropic.WithTokenSource(func(ctx context.Context) (string, error) {
            tok, err := creds.TokenSource.Token()
            if err != nil {
                return "", err
            }
            return tok.AccessToken, nil
        }),
    )
    if err != nil {
        log.Fatal(err)
    }


    stream, err := p.CompletionStream(ctx, llmrouter.ChatRequest{
        // Vertex addresses Claude with an @-versioned model id.
        Model:     "claude-3-5-sonnet-v2@20241022",
        MaxTokens: 1024,
        Messages: []llmrouter.Message{
            llmrouter.TextMessage("user", "In one sentence, what is Vertex Model Garden?"),
        },
    })
    if err != nil {
        log.Fatal(err)
    }


    for chunk := range stream.Chunks() {
        for _, c := range chunk.Choices {
            fmt.Print(c.Delta.Content)
        }
    }
    if err := stream.Err(); err != nil {
        log.Fatal(err)
    }
    fmt.Println()
}

Supported regions

Vertex serves Anthropic models from a subset of regions. Check the current list at cloud.google.com/vertex-ai/generative-ai/docs/partner-models/claude. At time of writing, us-east5 and europe-west1 are the most widely available regions for the current Claude 3.5 generation.

Error handling

Non-2xx responses surface as *llmrouter.ErrUpstream with Provider == "vertexanthropic". Common cases:

  • 401 — access token expired or wrong scopes (need cloud-platform).
  • 403 — IAM permission missing on the project; the caller needs aiplatform.endpoints.predict or the broader roles/aiplatform.user.
  • 404 — model id not available in this region (check the supported-regions matrix).
  • 400 — body shape mismatch; the most common cause is sending a model id without the @-version suffix.

When to use this vs other Anthropic surfaces

  • Use vertexanthropic when you're on GCP and want IAM + project + region as the auth and billing boundary, or your data-residency requirements pin you to a GCP region.
  • Use direct Anthropic when you want the latest variants on day one. Vertex Model Garden lags direct Anthropic by days to weeks for new releases.
  • Use Bedrock when you're on AWS — same shape, different cloud.
  • Use azureanthropic when you're on Azure.

See also