Built-In Google Cloud Vertex AI: Codex CLI Adds a Native GCP Provider

Sketchnote diagram for: Built-In Google Cloud Vertex AI: Codex CLI Adds a Native GCP Provider

Release track: v0.164.0-alpha.11.1 (23 October 2026, 01:58 UTC) Key PR: #56265 Impact: Google Cloud users, enterprise GCP deployments, multi-cloud Codex workflows

A single PR landed quietly in a hotfix branch on 23 October, shortly before the v0.164.0 stable release: Codex CLI now ships a built-in Google Cloud Vertex AI provider. Set model_provider = "gcp" in your configuration and Codex will route inference through Vertex AI without any custom HTTP adapter or third-party wrapper.

The feature did not make the v0.164.0 stable cut — the hotfix was branched off alpha.11 after the alpha.12 main branch had already frozen for stable — but it is merged and tested, and is expected to ship in v0.165.0.

How it works

Configuration

Add a single line to your Codex config to switch the active provider:

model_provider = "gcp"

Codex derives the Vertex AI endpoint from two environment variables:

Variable	Purpose	Default
`GOOGLE_CLOUD_PROJECT`	GCP project ID	required
`GOOGLE_CLOUD_LOCATION`	Vertex AI region	`"global"`

The global location default means you can get started with only a project ID set.

Authentication

Credentials are obtained by shelling out to the gcloud CLI:

gcloud auth application-default print-access-token

This follows Application Default Credentials conventions familiar to any GCP practitioner. If you need finer control, the provider supports three overrides under model_providers.gcp:

base_url — override the derived Vertex endpoint entirely
auth — supply a custom auth mechanism
http_headers — inject additional request headers

Model catalogue

Codex ships a static model catalogue tailored to Vertex AI when the GCP provider is active. The catalogue filters out models that Vertex does not expose and surfaces the Vertex-specific model identifiers you would use in the Google Cloud console.

Standard Codex features that work with the GCP provider:

Function tools
Tool namespacing
Client-side tool search
Remote compaction V2

Features explicitly disabled (Vertex AI does not support them):

Reasoning summaries
Responses Lite
Custom tools (server-side)
Web search
Image generation

Account reporting

The app-server protocol gains a new "gcp" account type. This means the Codex UI can display your GCP project context alongside OpenAI and AWS Bedrock accounts — account reads do not require OpenAI authentication and credentials are not resolved at account-read time.

Why this matters

Codex CLI has been OpenAI-first by design, with Bedrock arriving as the first major alternative provider. The GCP Vertex AI provider completes the major cloud triad:

Provider	Config key	Auth mechanism
OpenAI	`openai` (default)	API key / OAuth
Amazon Bedrock	`bedrock`	AWS credentials / wizard
Google Cloud Vertex AI	`gcp`	Application Default Credentials

For organisations already invested in Google Cloud — Vertex AI model deployments, Cloud Run workloads, BigQuery pipelines — this removes the last friction point for adopting Codex CLI as an agentic layer. No custom provider configuration, no HTTP shim: just a project ID and a gcloud login.

What to expect next

Because this PR landed on an alpha.11 hotfix branch rather than the main development line, it will need to be merged or cherry-picked into the v0.165.0 alpha sprint before shipping in a stable release. Watch for model_provider = "gcp" to appear in the v0.165.0 release notes.

If you want to try it now, build from the alpha.11.1 tag or wait for v0.165.0-alpha.1.

Source: PR #56265 — merged into v0.164.0-alpha.11.1 (23 October 2026, 01:58 UTC).