@fullsend-ai/pi-anthropic-vertex

September 4, 2026 · View on GitHub

Anthropic Claude models on Google Cloud Vertex AI, as a pi provider.

Registers the provider anthropic-vertex, with pi's whole first-party Claude catalog behind it: anthropic-vertex/claude-sonnet-5, anthropic-vertex/claude-opus-5, and so on.

1. Install

pi install git:github.com/fullsend-ai/pi-anthropic-vertex

2. Configure

Two things: a GCP project, and Application Default Credentials.

export GOOGLE_CLOUD_PROJECT=your-gcp-project
export CLOUD_ML_REGION=us-east5          # optional; us-east5 is the default
gcloud auth application-default login

No ANTHROPIC_API_KEY, and no pi login. Auth is ambient: the provider mints a short-lived Google access token from ADC on each request, and google-auth-library caches it and re-mints it near expiry — so a fresh machine or a CI sandbox needs nothing beyond ADC.

	Variable	Default
Project	`GOOGLE_CLOUD_PROJECT`, then `GCLOUD_PROJECT`, then `ANTHROPIC_VERTEX_PROJECT_ID`, then `GOOGLE_CLOUD_PROJECT_ID` — first non-empty wins	none; without one the provider stays out of the model list
Region	`CLOUD_ML_REGION`, then `GOOGLE_CLOUD_LOCATION`	`us-east5`
Credentials	`GOOGLE_APPLICATION_CREDENTIALS`, or whatever else ADC discovers	ADC

region also accepts Vertex's non-regional endpoints: global, us, and eu (the data-residency endpoints) map to their own hostnames.

No other ANTHROPIC_* variable is read anywhere — not ANTHROPIC_API_KEY, not ANTHROPIC_VERTEX_BASE_URL. If your launcher unsets ANTHROPIC_* before starting pi (fullsend does), that changes nothing here.

3. Check that it loaded

A pi extension that fails to load is dropped silently, so confirm the provider is there before debugging anything else:

$ pi --list-models | grep -E 'provider|anthropic-vertex'
provider          model                       context  max-out  thinking  images
anthropic-vertex  claude-fable-5              1M       128K     yes       yes
anthropic-vertex  claude-fable-5-1            1M       128K     yes       yes
anthropic-vertex  claude-haiku-4-5            200K     64K      yes       yes
anthropic-vertex  claude-haiku-4-5-20251001   200K     64K      yes       yes
anthropic-vertex  claude-opus-4-5             200K     64K      yes       yes
anthropic-vertex  claude-opus-4-5-20251101    200K     64K      yes       yes
anthropic-vertex  claude-opus-4-6             1M       128K     yes       yes
anthropic-vertex  claude-opus-4-7             1M       128K     yes       yes
anthropic-vertex  claude-opus-4-8             1M       128K     yes       yes
anthropic-vertex  claude-opus-5               1M       128K     yes       yes
anthropic-vertex  claude-sonnet-4-5           1M       64K      yes       yes
anthropic-vertex  claude-sonnet-4-5-20250929  1M       64K      yes       yes
anthropic-vertex  claude-sonnet-4-6           1M       128K     yes       yes
anthropic-vertex  claude-sonnet-5             1M       128K     yes       yes

(That listing is pi 0.85.0's catalog; yours reflects whatever pi you are running — see Models.)

4. Run it

pi --model anthropic-vertex/claude-sonnet-5 "explain this repo in three sentences"

Verified 2026-09-04 with pi 0.84.4, ADC from gcloud auth application-default login, and CLOUD_ML_REGION=global, against a project where these models are enabled — the current default Sonnet, a --thinking medium call on an adaptive model, and a dated id (rewritten to claude-haiku-4-5@20251001 on the wire):

$ pi -ne -e . --model anthropic-vertex/claude-sonnet-5 --thinking off --no-session -p "reply with just: ok"
ok
$ pi -ne -e . --model anthropic-vertex/claude-sonnet-4-6 --thinking medium --no-session -p "reply with just: ok"
ok
$ pi -ne -e . --model anthropic-vertex/claude-haiku-4-5-20251001 --thinking off --no-session -p "reply with just: ok"
ok

(-ne -e . loads only this extension from the current checkout, the way the CI smoke test and fullsend's sandbox do; a normal pi install needs neither flag.)

Or set it as the default in ~/.pi/agent/settings.json:

{ "defaultProvider": "anthropic-vertex", "defaultModel": "claude-sonnet-5" }

Use the fully qualified anthropic-vertex/claude-sonnet-5 in scripts. The bare id resolves to pi's built-in anthropic provider, which wants an ANTHROPIC_API_KEY and bills through api.anthropic.com.

Models

The catalog is not maintained here. It is read at load time from pi's own first-party Anthropic catalog, so:

Model ids are pi's ids — claude-sonnet-5, claude-opus-5, claude-opus-4-8, claude-opus-4-6, claude-fable-5-1, claude-haiku-4-5. Your model specs and any alias table keep working across this extension and the built-in anthropic provider.
Pricing, context windows and thinking levels are pi's, copied verbatim, so cost reporting matches the direct-Anthropic path exactly.
A pi upgrade brings new models with it. Nothing to release here when Anthropic ships one.

Two details are handled on the wire and never appear in a model spec:

Vertex names dated models with @ — claude-haiku-4-5-20251001 is claude-haiku-4-5@20251001 in the URL. The rewrite does that; the id you type does not change.
Anthropic capability flags that Vertex does not accept are filtered out (see Troubleshooting and CONTRIBUTING.md).

Whether a given model is served to your project is a Vertex question, not a pi one: a model can be listed here and still 404 in your region.

In fullsend

fullsend's pi runtime sets the project and region for you and expects this provider id. Per-agent selection is an agents: entry in .fullsend/config.yaml:

agents:
  - name: triage
    runtime: pi
    model: anthropic-vertex/claude-opus-5

or:

fullsend agent set triage --fullsend-dir .fullsend --runtime pi --model anthropic-vertex/claude-opus-5

The provider id, the environment contract and the disabled message below are all byte-compatible with twoGiants/pi-anthropic-vertex, so this is a drop-in replacement for it in a sandbox image. See the pi runtime docs for the full precedence rules.

For Grok on Vertex under the same ADC login, see the sibling extension fullsend-ai/pi-xai-vertex. One gcloud auth application-default login covers every Vertex provider — it is the project that differs between them, not the identity.

Requirements

pi ≥ 0.84.4 (tested against 0.84.4 and 0.85.0 on every commit)
A GCP project with the Claude models enabled in Vertex AI Model Garden
ADC credentials (gcloud auth application-default login, or GOOGLE_APPLICATION_CREDENTIALS)

Troubleshooting

[pi-anthropic-vertex] disabled: set GOOGLE_CLOUD_PROJECT or ANTHROPIC_VERTEX_PROJECT_ID. No project resolved, so nothing was registered. Set one of the four project variables above. This is the only case where the extension deliberately does not register.

The provider is missing from --list-models and there is no message. pi drops a failed extension without printing anything. Load it explicitly to see the real error:

pi -e ~/.pi/agent/git/github.com/fullsend-ai/pi-anthropic-vertex --list-models

[anthropic-vertex] Google ADC unavailable, .... ADC could not be discovered, so the models are registered but not offered. Run gcloud auth application-default login, or point GOOGLE_APPLICATION_CREDENTIALS at a credential config. The message names the underlying cause and is printed once per process.

No API key found for anthropic. You used a bare model id and reached pi's built-in provider. Use the fully qualified anthropic-vertex/<model>.

A 404 on a model that appears in --list-models. Vertex does not serve that model to your project in that region. Try another region (CLOUD_ML_REGION=us-east5, europe-west1, global) or enable the model in Model Garden. The catalog comes from pi and cannot know your project's grants.

A 400 naming a request field — fallbacks: Extra inputs are not permitted, or an unrecognised beta. pi has started sending a field Vertex does not accept. Anthropic capability flags reach Vertex through an explicit allowlist (VERTEX_COMPAT_KEYS in src/index.ts) precisely so this stays rare; the fix is to leave the new flag out of that list, with a note saying why. See CONTRIBUTING.md.

Contributing, architecture, and the pi-specific gotchas behind the design: CONTRIBUTING.md. Rules for AI agents working in this repo: AGENTS.md.

MIT licensed.