Nexura Vertex

Google Vertex AI models for Nexura agents, using Google Application Default Credentials.

Install

pip install nexura-vertex

Alpha (0.4): pip install nexura-vertex. See package installation.

Minimal example

This example uses mock HTTP and a fixed fake token, without credentials or calls to Google. For live calls, use Agent(model="vertex:gemini-2.5-flash") after configuring ADC (gcloud auth application-default login, a service account or VM metadata). Project selection also accepts GOOGLE_CLOUD_PROJECT or project=; custom tokens use token_provider=. Live calls are billed to your project.

import httpx
from nexura_agents import Agent, Runner

from nexura_vertex import VertexAIProvider


def fake_vertex(request: httpx.Request) -> httpx.Response:
    assert request.headers["Authorization"] == "Bearer test-token"
    assert "/projects/demo/locations/global/" in request.url.path
    return httpx.Response(
        200,
        json={
            "candidates": [
                {"content": {"role": "model", "parts": [{"text": "Hello"}]}, "finishReason": "STOP"}
            ],
            "usageMetadata": {"promptTokenCount": 3, "candidatesTokenCount": 1},
        },
    )


model = VertexAIProvider(
    "gemini-2.5-flash",
    project="demo",
    token_provider=lambda: "test-token",
    transport=httpx.MockTransport(fake_vertex),
)
result = Runner().run(Agent(name="Hi", instructions="Greet.", model=model), "hi")
print(result.final_output)

Status and limits

Status: alpha (0.4). Gemini 2.5 Flash on Vertex AI passed one live check run by the repository owner on 2026-10-08 (a plain reply, a tool call and streaming). Everything else is tested offline with mock HTTP. Depends only on nexura-agents and google-auth, not nexura-runtime; Python 3.11–3.13. The adapter reuses Gemini tools/streaming/structured output/multimodal parts, with OAuth endpoints and :predict embeddings.

Documentation and licence

Main documentation · Full guide. Apache-2.0, Copyright (c) 2026 Liceron. See LICENSE.