GCP AI — Google Cloud credits and Vertex models (Magpie / OpenCode plugin)

• Models: the Vertex chat models your project can call, newest first.
• Inference: requests speak Gemini's own protocol (@ai-sdk/google) to the project's Vertex endpoint, signed with your local gcloud / ADC credentials and billed to your own project.
• Credits: the quota panel can show your Google Cloud promotional credit balance and its expiry date.

You sign in once with your own Google account. No service accounts, no JSON keys, no extra packages: the plugin uses the Chrome you already have.

Requirements

• Google Chrome or Chromium (set VERTEX_CHROME_PATH if it is somewhere unusual)
• Google Cloud CLI (gcloud), already working — on Windows the gcloud.cmd the installer puts on PATH is used
• A Google Cloud project with Vertex AI enabled
• To read credits: permission to view the project's billing account (usually billing.viewer)

Linux has been tried with real accounts. Support for macOS and Windows is implemented, but has not been tried with a real account on either platform. On Windows the plugin's own files live under %APPDATA% and %LOCALAPPDATA% instead of ~/.config and ~/.cache, and gcloud is started through cmd.exe, which is the only way Windows runs a .cmd.

What this was run on

Sign-in, model listing and a real inference request were run on Linux against two real Google accounts, with gcloud's own ADC and with an ADC the plugin wrote.

The Windows path — cmd.exe for gcloud.cmd, %APPDATA%/%LOCALAPPDATA% for the plugin's own files — is implemented and covered by the unit tests' path handling, but has not been run on a Windows machine with a real account. If it misbehaves there, that is why; macOS has likewise not been tried with a real account.

Install and sign in

magpie plugin add @magpie-community/gcp-ai

Then click sign in on the GCP AI row, or run:

magpie plugin login gcp-ai

One Chrome window opens. Approve the consent page there and the plugin closes the window itself — never close it yourself.

• The first account goes straight to Google's consent page. Adding a second one starts on Google's account picker instead, in the same window.
• Add another account by clicking sign in again with a different Google account.
• The project is picked for you: the first of the account's projects with the Vertex AI API enabled.
• If the account already has an ADC here, nothing is added and your existing sign-in is left alone.

Credits

Google has no API for promotional credits, so the balance is read from the Cloud Console page. The row opens a browser when it needs a balance, and only then. Balances are cached for five minutes.

What it keeps where

• Credentials: never held by the plugin. gcloud writes them and prints tokens on demand.
• ADC directories: <config>/gcp-ai/adc-<account>, one per Google account. The plugin never writes gcloud's own directory. Two addresses that differ only in punctuation would share a directory name, so the second one gets a short hash of its address appended.
• Chrome profile: <config>/gcp-auto-profile.
• Cache: <cache>/magpie/gcp-ai.

<config> is $XDG_CONFIG_HOME, else %APPDATA% on Windows, else ~/.config. <cache> is $XDG_CACHE_HOME, else %LOCALAPPDATA% on Windows, else ~/.cache.

FAQ

No credits shown? The Chrome sign-in has probably expired. Sign in again.

"This browser or app may not be secure"? Start the sign-in again from the row and use the window that opens. Nothing in it is automated.

"That account was already signed in"? Signing in adds an account; it does not refresh one. Use a different Google account.

Wrong account on the consent page? Sign in again and pick the right one. Nothing that was already here is removed.

Credits don't match the billing page? Promotional credits belong to the billing account and can be shared across projects, and the Console has billing latency.

Configuration

Environment variable	Default
GOOGLE_VERTEX_LOCATION	global
VERTEX_BILLING_ACCOUNT	Looked up from the project
VERTEX_CHROME_PROFILE	~/.config/gcp-auto-profile
VERTEX_CHROME_PATH	Detected
VERTEX_GCLOUD_ACCOUNT	The signed-in account
VERTEX_GCLOUD_AUTH_MODE	auto; also adc or gcloud
VERTEX_CHROME_IDLE	30 seconds a kept browser waits before closing; 0 closes it right away
VERTEX_CHROME_GAP	0 seconds between one process closing Chrome and the next opening the profile
VERTEX_CREDITS_CACHE_TTL	300 seconds
VERTEX_CREDITS_FAILURE_TTL	1200 seconds
VERTEX_CREDITS_TRANSIENT_TTL	60 seconds, for network failures
VERTEX_MODELS_CACHE_TTL	21600 seconds (6 hours)

A machine-wide CLOUDSDK_CONFIG is never read. It would point every account at one ADC, and a sign-in there would overwrite credentials the plugin didn't write; the plugin only uses its own <config>/gcp-ai/adc-<account> directories.

Privacy

The plugin asks Google for your model list with your credentials, sends inference requests to the Vertex endpoint, and opens Cloud Console in a Chrome it started to read your credits. No third-party server, no credential upload, no debugging port.