Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

docs(vertex): how to point Jarvis at a Google Cloud project

Haraldecdafbf7bfac
diff --git a/.env.example b/.env.example
index 6c21b9a4bd..6a001376f5 100644
--- a/.env.example
+++ b/.env.example
@@ -21,6 +21,19 @@
# Gemini (Google AI Studio) — default Brain. https://aistudio.google.com/app/apikey
GEMINI_API_KEY=
GOOGLE_AIStudio_API_KEY=
+# Google Cloud Vertex AI — the SAME Gemini models, billed to a Cloud project
+# instead of an AI Studio account. https://console.cloud.google.com/vertex-ai/studio
+# One credential serves brain, tool model, STT, TTS, realtime and subagents.
+# Either an express-mode key (AQ.) or a Cloud API key restricted to
+# aiplatform.googleapis.com. For a production project set no key at all and use
+# JARVIS__GOOGLE__VERTEX_PROJECT below (Application Default Credentials).
+VERTEX_API_KEY=
+JARVIS__GOOGLE__VERTEX_PROJECT=
+# Region for the project path. "global" has the widest Gemini availability;
+# name a single region when data residency requires it.
+JARVIS__GOOGLE__VERTEX_LOCATION=
+# Only needed when this machine is not already signed in with gcloud.
+JARVIS__GOOGLE__SERVICE_ACCOUNT_PATH=
# Anthropic Claude. https://console.anthropic.com/settings/keys
ANTHROPIC_API_KEY=
# OpenAI (GPT brain, Whisper STT, TTS). https://platform.openai.com/api-keys
diff --git a/CHANGELOG.md b/CHANGELOG.md
index c137ca39de..20bf13c4e2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,48 @@ versioning per [SemVer](https://semver.org/).
## [Unreleased]
+### Added
+
+- **Google Cloud Vertex AI as a provider family of its own, on every tier.**
+ Vertex serves the same Gemini models as Google AI Studio, but bills a Cloud
+ project — and until now it was reachable only by accident: an express-mode key
+ pasted into the Gemini field, detected by a probe, sharing that field so an
+ install could never hold both credentials. Vertex now has its own cards for
+ the brain, the tool model, voice input, voice output and realtime, and is
+ selectable as a subagent. All of them read ONE shared Vertex credential, so
+ setting it up once covers the whole stack.
+ - Two ways in. A Vertex AI API key — express mode (`AQ.`) or a Cloud API key
+ restricted to `aiplatform.googleapis.com` (`AIza`) — or a Google Cloud
+ project with no key at all: `[google].vertex_project` plus
+ `vertex_location` and an optional `service_account_path`, signed with
+ Application Default Credentials. That second path is what a production
+ project uses, and it was previously wired to text-to-speech only.
+ - The endpoint is now decided by the card, not guessed from the key. A Cloud
+ API key restricted to Vertex wears the ordinary `AIza` shape, so the
+ existing routing probe sent it to AI Studio, where it failed every call
+ with an auth error that explained nothing. Keys in a Vertex card are pinned
+ to Vertex with no probe at all.
+ - Vertex is a distinct credential family, deliberately with no cross-read to
+ the Gemini slots in either direction. The two accounts run out of quota
+ independently, which makes crossing between them a real fallback rather
+ than a doomed retry — and an AI Studio key forced onto the Vertex endpoint
+ would not degrade, it would 401 while the UI showed a configured provider.
+ - Worth picking for voice: the AI Studio text-to-speech preview model is
+ capped at 100 requests per day no matter how you are billed, which is what
+ made the voice go quiet mid-day. A Cloud project has no such cap.
+
+- **Gemini speech-to-text got the card it never had.** The recognizer shipped as
+ a working entry point with no card anywhere in the app, so selecting it meant
+ hand-editing `jarvis.toml`. Both Google recognizers now appear in Voice Input
+ with a model picker.
+
+### Fixed
+
+- The key-format hint no longer calls an `AIza` key a "Google AI Studio key".
+ The same shape is issued for Google Cloud, so the label is endpoint-neutral
+ now and the note says which card decides where the key is sent. A correctly
+ pasted Vertex key also gets its green confirmation instead of silence.
+
---
## [1.3.2] — 2026-08-12
diff --git a/docs/product/personalize-and-connect/providers-and-api-keys.md b/docs/product/personalize-and-connect/providers-and-api-keys.md
index fda92ab090..ca31da5946 100644
--- a/docs/product/personalize-and-connect/providers-and-api-keys.md
+++ b/docs/product/personalize-and-connect/providers-and-api-keys.md
@@ -77,6 +77,37 @@ Card labels describe state, not a guarantee:
with `AQ.`) into the same field — the app detects which service issued it
and routes every capability (brain, tool model, realtime, speech) through
the matching endpoint automatically.
+
+### Google Cloud Vertex AI
+
+Google serves the same Gemini models twice: through an AI Studio account, and
+through Google Cloud Vertex AI. There are separate **Vertex AI** cards for the
+brain, voice input, voice output, and realtime, plus Vertex as a subagent, so
+one Vertex setup covers the whole stack — and so an install can hold an AI
+Studio key and a Vertex credential side by side and point each capability at
+the one it should bill.
+
+Two ways to authenticate, and you only need one:
+
+- **A Vertex AI API key.** Either an express-mode key (starts with `AQ.`) or a
+ Google Cloud API key restricted to `aiplatform.googleapis.com` (starts with
+ `AIza`). Paste it into any Vertex card; the cards share one credential.
+- **A Google Cloud project, with no key at all.** Set `[google]` in
+ `jarvis.toml`: `vertex_project` to the project ID that owns the Vertex AI
+ API, `vertex_location` to a region (`global` is the default and has the
+ widest model availability), and `service_account_path` only if this machine
+ is not already signed in with `gcloud`. Requests are then signed with
+ Application Default Credentials. This is the path a production project uses.
+
+Two reasons to prefer Vertex over the AI Studio cards: the billing lands in
+your Cloud project, and the text-to-speech preview model that AI Studio caps at
+100 requests per day — regardless of how you are billed — has no such cap
+there. That cap is what used to make the voice go quiet mid-day.
+
+Because a Vertex key and an AI Studio key can look identical, the endpoint is
+decided by the card you save the key under, not by the key's shape. A key in a
+Vertex card always goes to Vertex; a key in a Gemini card is detected and
+routed as described above.
3. Paste the value into the password field and select **Save**. The field then
becomes masked. The page receives only whether the credential exists; it
does not read the saved value back.
diff --git a/jarvis.toml.example b/jarvis.toml.example
index d9414d67a3..a3d8d0672a 100644
--- a/jarvis.toml.example
+++ b/jarvis.toml.example
@@ -44,7 +44,7 @@ sensitivity = 0.5 # 0.5..1 (values below 0.5 are lifted — the wake
# ── Speech-to-Text ──
[stt]
-provider = "groq-api" # groq-api | openai-api | openrouter-stt | gemini-api | faster-whisper (local)
+provider = "groq-api" # groq-api | openai-api | openrouter-stt | gemini-api | vertex-stt | faster-whisper (local)
device = "cpu" # cpu | cuda — keep "cpu" unless you have a local GPU
language = "auto" # auto, or a weak ISO-639 preference such as "fr"
bias_prompt = "" # optional comma-separated vocab to stabilize proper nouns
@@ -80,16 +80,32 @@ voice_de = "Charon"
voice_en = "Charon"
language_code = "auto"
streaming = true
-# Vertex AI routing avoids the AI-Studio preview RPD cap. Off by default;
-# set JARVIS__TTS__VERTEX_PROJECT (env) + the SA path to enable.
+# Legacy TTS-only Vertex route (predates the [google] section below). Kept for
+# existing installs; new setups should pick provider = "vertex-tts" instead, which
+# shares ONE Vertex credential with the brain, STT and realtime tiers.
use_vertex = false
vertex_project = ""
service_account_path = "~/.config/jarvis/vertex-sa.json"
+# ── Google endpoint routing (AI Studio vs Google Cloud Vertex AI) ──
+# vertex_mode decides where an AMBIGUOUS key stored in a Gemini slot goes:
+# "auto" probe once per process and remember (default)
+# "always" force Vertex | "never" force AI Studio
+# The vertex_* fields below configure the dedicated Vertex provider family
+# (brain "vertex", stt "vertex-stt", tts "vertex-tts", realtime "vertex-live",
+# and the same slug as a subagent). Leave vertex_project empty to authenticate
+# with a Vertex API key; set it to use a Google Cloud project with Application
+# Default Credentials and no key at all.
+[google]
+vertex_mode = "auto"
+vertex_project = ""
+vertex_location = "global" # "global" = widest Gemini availability; or e.g. "europe-west4"
+service_account_path = "" # only when this machine is not signed in with gcloud
+
# ── Brain (multi-provider; never hardcode a single vendor) ──
# Cloud-reachable default = Gemini (needs only GEMINI_API_KEY). Switch the
# whole stack by changing these four to "claude-api" | "openai" | "grok" |
-# "openrouter" and supplying that provider's key.
+# "openrouter" | "vertex" and supplying that provider's key.
[brain]
primary = "gemini"
fallback = "gemini"