Feature request: Add native Google Vertex AI provider for Claude models

Feature request: Add native Google Vertex AI provider for Claude models

Problem

Nanobot currently ships providers for Anthropic (direct), OpenAI, Azure OpenAI, AWS Bedrock, GitHub Copilot, xAI, and various OpenAI-compat gateways, but no first-class provider for Anthropic Claude models running on Google Vertex AI.

Enterprise/GCP-native users who consume Claude through Vertex (rather than the Anthropic API directly or AWS Bedrock) currently have no supported path in nanobot.

What's missing vs. what already exists

I looked through the repo and prior work before filing this:

• nanobot/providers/ has anthropic_provider.py and bedrock_provider.py but no vertex_ai_provider.py.
• Existing open PRs #1943 and #2377 both add Vertex AI support for Gemini models (via LiteLLM / google-cloud-aiplatform). Neither adds a path for Claude on Vertex.
• Previous closed PRs (#111, #131, #1324) and closed issue #2082 also focused on Gemini or were closed as mistakes.
• So this is a distinct feature: Claude models on Vertex AI via Anthropic's own SDK, not Gemini via Google's SDK.

Proposed solution

Add a VertexAIProvider that subclasses AnthropicProvider and uses the official AsyncAnthropicVertex client from the anthropic Python package:

from anthropic import AsyncAnthropicVertex

client = AsyncAnthropicVertex(
    project_id=project,   # ANTHROPIC_VERTEX_PROJECT_ID or GOOGLE_CLOUD_PROJECT
    region=region,        # CLOUD_ML_REGION or GOOGLE_CLOUD_LOCATION, default us-central1
    max_retries=0,
)

This is a very small change because:

• The anthropic SDK already exposes AsyncAnthropicVertex with the same interface as AsyncAnthropic, so almost all of AnthropicProvider's logic (streaming, tool_use, prompt caching, etc.) is reusable via subclassing.
• Authentication uses Google Application Default Credentials (ADC) — obtained via gcloud auth application-default login — so no API key needs to be stored in config.

Sketch of the changes

1. nanobot/providers/vertex_ai_provider.py (new, ~120 lines)

   • VertexAIProvider(AnthropicProvider) that instantiates AsyncAnthropicVertex in __init__.
   • Strips provider prefixes like google_vertex_ai/, vertex/, anthropic/ from the model string.
   • Helpers: get_adc_status(), login_vertex_ai() (runs gcloud auth application-default login), logout_vertex_ai() (runs gcloud auth application-default revoke).

2. nanobot/providers/factory.py — new branch:

   elif backend == "google_vertex_ai":
       from nanobot.providers.vertex_ai_provider import VertexAIProvider
       provider = VertexAIProvider(
           default_model=model,
           project=getattr(p, "project", None) if p else None,
           region=getattr(p, "region", None) if p else None,
       )
   

3. nanobot/providers/registry.py — new ProviderSpec:

   ProviderSpec(
       name="google_vertex_ai",
       keywords=("vertex", "vertex_ai", "google_vertex"),
       env_key="",
       display_name="Google Vertex AI",
       backend="google_vertex_ai",
       is_oauth=True,
       strip_model_prefix=True,
   ),
   

4. config/schema.py — add optional project and region fields under providers.googleVertexAi.

5. Optional dependency: pip install nanobot-ai[vertex] → pulls in google-auth (already a transitive dep of the anthropic SDK's Vertex extras).

Example config

providers:
  googleVertexAi:
    project: my-gcp-project
    region: us-east5
model: google_vertex_ai/claude-sonnet-4-5@20250929

Why do this in addition to PRs #1943 and #2377

Those PRs address Gemini on Vertex. Claude on Vertex is a common enterprise pattern (Anthropic Claude models are GA on Vertex in us-central1, us-east5, europe-west1, etc.) and can't be served by the Gemini-focused providers because the on-the-wire API is Anthropic's, not Google's — hence needing AsyncAnthropicVertex.

Ideally the two efforts land as sibling providers under the same google_vertex_ai umbrella, with the model prefix (or an explicit provider setting) selecting the correct backend client. Happy to coordinate with the authors of #1943 / #2377.

Prior art / working implementation

I have this working locally in a downstream fork against main. Total change is ~200 lines across the four files listed above, with no changes required to the base AnthropicProvider. Happy to open a PR if there's interest — please let me know before I put in the effort so we don't add to the pile of stalled Vertex PRs.

Environment

• nanobot version: latest main
• Python: 3.11
• anthropic>=0.39.0 (any version exposing AsyncAnthropicVertex)
• Auth: Google Application Default Credentials via gcloud auth application-default login