--- summary: "Google Gemini setup (AI Studio API key, Vertex AI, optional CLI runtime, and multimodal tools)" title: "Google (Gemini)" read_when: - You want to use Google Gemini models with OpenClaw - You need Google AI Studio, Vertex AI, or Gemini CLI runtime guidance --- The Google plugin provides access to Gemini models through Google AI Studio, plus image generation, media understanding (image/audio/video), text-to-speech, and web search via Gemini Grounding. - Provider: `google` - Auth: `GEMINI_API_KEY` or `GOOGLE_API_KEY` - API: Google Gemini API - Managed-cloud provider: `google-vertex` with Google Cloud Application Default Credentials - Optional runtime: `agentRuntime.id: "google-gemini-cli"` runs an explicitly configured model through the local Gemini CLI ## Getting started For most installations, use a Google AI Studio API key. Use `google-vertex` when the Gateway already runs inside a managed Google Cloud environment. **Recommended for:** standard Gemini API access. Create a free key in [Google AI Studio](https://aistudio.google.com/apikey). ```bash openclaw onboard --auth-choice gemini-api-key ``` Or pass the key directly: ```bash openclaw onboard --non-interactive --accept-risk --skip-health \ --mode local \ --auth-choice gemini-api-key \ --gemini-api-key "$GEMINI_API_KEY" ``` ```json5 { agents: { defaults: { model: { primary: "google/gemini-3.1-pro-preview" }, }, }, } ``` ```bash openclaw models list --provider google ``` `GEMINI_API_KEY` and `GOOGLE_API_KEY` are both accepted. Use whichever you already have configured. With a configured API key, OpenClaw refreshes Google AI Studio's text-model catalog from the Gemini `models.list` API. Newly released Gemini 3 Pro, Flash, and Flash-Lite variants therefore appear in `openclaw models list --provider google` without waiting for an OpenClaw release. Failed refreshes report the failure and retain the last successful inventory, or bundled models before the first success. A successful empty response clears discovered models. Vertex uses its separate static catalog. **Recommended for:** Gateway instances running inside Google Cloud (GCE, GKE, Cloud Run) or workstations already authenticated through gcloud Application Default Credentials. On a workstation, run: ```bash gcloud auth application-default login ``` On a Google Cloud VM with an attached service account, the metadata server already provides credentials and no file is needed. Both are required in the gateway service environment: ```bash export GOOGLE_CLOUD_PROJECT="my-project" export GOOGLE_CLOUD_LOCATION="us-central1" ``` `GCLOUD_PROJECT` is also accepted for the project. Without a project, runs fail with `Vertex AI requires a project ID. Set GOOGLE_CLOUD_PROJECT/GCLOUD_PROJECT or pass project in options.` Without a location, the error is `Vertex AI requires a location. Set GOOGLE_CLOUD_LOCATION or pass location in options.` With a recognized ADC file (created by `gcloud auth application-default login`) **and** the project and location variables set, OpenClaw discovers `google-vertex` automatically — nothing needs to be stored. Manual storage is only needed for metadata-only credentials (for example a GCE VM with an attached service account and no ADC file on disk) or to select the ADC path explicitly. The ADC path is selected by a sentinel value, not a real key — store exactly `gcp-vertex-credentials`: ```bash openclaw models auth paste-api-key --provider google-vertex # paste exactly: gcp-vertex-credentials ``` Other stored values are sent in the `x-goog-api-key` header, not as ADC credentials. Do not paste an OAuth access token from `gcloud auth print-access-token`: it is not an API key and can cause an authentication error such as `401 UNAUTHENTICATED`. ```bash openclaw models list --refresh --all --provider google-vertex ``` `--all` includes models you have not configured. `--refresh` populates the local catalog when it has not been loaded yet; it does not query a live Vertex model inventory or verify that ADC can obtain a token. Vertex uses a separate static model catalog; new Gemini variants appear there per OpenClaw release rather than through the live refresh used by the AI Studio key path. **Advanced use only:** run a canonical `google/*` model through an installed Gemini CLI while keeping authentication on the supported AI Studio API-key path. OpenClaw does not offer new Gemini CLI OAuth or Antigravity OAuth setup. [Google ended consumer Gemini CLI Login with Google access on June 18, 2026](https://developers.google.com/gemini-code-assist/docs/deprecations/code-assist-individuals), and the [Antigravity terms](https://antigravity.google/terms) prohibit third-party tools from accessing the service through Antigravity OAuth. Use an AI Studio API key or Vertex AI instead. Complete the API-key setup in the first tab. OpenClaw must have a usable `google` API-key profile before the CLI runtime can be selected. The local `gemini` command must be available on `PATH`. ```bash # Homebrew brew install gemini-cli # or npm npm install -g @google/gemini-cli ``` OpenClaw supports both Homebrew installs and global npm installs, including common Windows/npm layouts. Keep the canonical Google model ref and opt that model into the CLI runtime: ```json5 { agents: { defaults: { model: { primary: "google/gemini-3.1-pro-preview" }, models: { "google/gemini-3.1-pro-preview": { agentRuntime: { id: "google-gemini-cli" }, }, }, }, }, } ``` - Runtime: `google-gemini-cli` - Auth: selected Google AI Studio API-key profile - Model refs: canonical `google/*` `google-gemini-cli` is the CLI backend the bundled Google plugin registers. See [CLI backends](/gateway/cli-backends) for its argv, JSONL dialect, and session and compaction behavior, and for the settings shared by every registered backend. Existing valid Gemini CLI OAuth profiles remain executable for compatibility, but OpenClaw cannot create or repair them. If one breaks, replace it with a Google AI Studio API-key profile. `google-gemini-cli/*` refs remain legacy compatibility aliases. New configs should use `google/*` model refs plus the explicit runtime selection above. `google/gemini-3-pro-preview` was retired on 2026-03-09; use `google/gemini-3.1-pro-preview` instead. Re-running Gemini API key setup (`openclaw onboard --auth-choice gemini-api-key` or `openclaw models auth login --provider google`) rewrites a stale configured default to the current model. ## Capabilities | Capability | Supported | | ---------------------- | ----------------------------- | | Chat completions | Yes | | Image generation | Yes | | Music generation | Yes | | Text-to-speech | Yes | | Realtime voice | Yes (Google Live API) | | Image understanding | Yes | | Audio transcription | Yes | | Video understanding | Yes | | Web search (Grounding) | Yes | | Thinking/reasoning | Yes (Gemini 2.5+ / Gemini 3+) | | Gemma 4 models | Yes | ## Web search The bundled `gemini` web-search provider uses Gemini Google Search grounding. Configure a dedicated search key under `plugins.entries.google.config.webSearch`, or let it reuse `models.providers.google.apiKey` after `GEMINI_API_KEY`: ```json5 { plugins: { entries: { google: { config: { webSearch: { apiKey: "AIza...", // optional if GEMINI_API_KEY or models.providers.google.apiKey is set baseUrl: "https://generativelanguage.googleapis.com/v1beta", // falls back to models.providers.google.baseUrl model: "gemini-3.6-flash", }, }, }, }, }, } ``` Credential precedence is dedicated `webSearch.apiKey`, then `GEMINI_API_KEY`, then `models.providers.google.apiKey`. `webSearch.baseUrl` is optional and exists for operator proxies or compatible Gemini API endpoints; when omitted, Gemini web search reuses `models.providers.google.baseUrl`. See [Gemini search](/tools/gemini-search) for the provider-specific tool behavior. Gemini 3 models use `thinkingLevel` rather than `thinkingBudget`. OpenClaw maps Gemini 3, Gemini 3.1, and `gemini-*-latest` alias reasoning controls to `thinkingLevel` so default/low-latency runs do not send disabled `thinkingBudget` values. `/think adaptive` keeps Google's dynamic thinking semantics instead of choosing a fixed OpenClaw level. Gemini 3 and Gemini 3.1 omit a fixed `thinkingLevel` so Google can choose the level; Gemini 2.5 sends Google's dynamic sentinel `thinkingBudget: -1`. Gemma 4 models (for example `gemma-4-26b-a4b-it`) support thinking mode. OpenClaw rewrites `thinkingBudget` to a supported Google `thinkingLevel` for Gemma 4. Setting thinking to `off` preserves thinking disabled instead of mapping to `MINIMAL`. Gemini 2.5 Pro only works in thinking mode and rejects an explicit `thinkingBudget: 0`; OpenClaw strips that value for Gemini 2.5 Pro requests instead of sending it. ## Image generation The bundled `google` image-generation provider defaults to `google/gemini-3.1-flash-image`. - Also supports `google/gemini-3-pro-image` - Generate: up to 4 images per request - Edit mode: enabled, up to 5 input images - Geometry controls: `size`, `aspectRatio`, and `resolution` To use Google as the default image provider: ```json5 { agents: { defaults: { mediaModels: { image: { primary: "google/gemini-3.1-flash-image", }, }, }, }, } ``` See [Image Generation](/tools/image-generation) for shared tool parameters, provider selection, and failover behavior. ## Video generation The bundled `google` plugin also registers video generation through the shared `video_generate` tool. - Default video model: `google/veo-3.1-fast-generate-preview` - Modes: text-to-video, image-to-video, and single-video reference flows - Supports `aspectRatio` (`16:9`, `9:16`) and `resolution` (`720P`, `1080P`); audio output is not supported by Veo today - Supported durations: **4, 6, or 8 seconds** (other values snap to the nearest allowed value) To use Google as the default video provider: ```json5 { agents: { defaults: { mediaModels: { video: { primary: "google/veo-3.1-fast-generate-preview", }, }, }, }, } ``` See [Video Generation](/tools/video-generation) for shared tool parameters, provider selection, and failover behavior. ## Music generation The bundled `google` plugin also registers music generation through the shared `music_generate` tool. - Default music model: `google/lyria-3-clip-preview` - Also supports `google/lyria-3-pro-preview` - Prompt controls: `lyrics` and `instrumental` - Output format: `mp3` by default, plus `wav` on `google/lyria-3-pro-preview` - Reference inputs: up to 10 images - Session-backed runs detach through the shared task/status flow, including `action: "status"` To use Google as the default music provider: ```json5 { agents: { defaults: { mediaModels: { music: { primary: "google/lyria-3-clip-preview", }, }, }, }, } ``` See [Music Generation](/tools/music-generation) for shared tool parameters, provider selection, and failover behavior. ## Text-to-speech The bundled `google` speech provider uses Gemini API TTS. The default model stays `gemini-3.1-flash-tts-preview`. Set `model` to `gemini-3.8-flash-tts` to opt in to Gemini 3.8, or `gemini-3.8-flash-lite-tts` for the faster, lower-cost variant. `gemini-2.5-flash-preview-tts` and `gemini-2.5-pro-preview-tts` remain available. - Default voice: `Kore` - Auth: `tts.providers.google.apiKey`, `models.providers.google.apiKey`, `GEMINI_API_KEY`, or `GOOGLE_API_KEY` - Output: WAV for regular TTS attachments, Opus for voice-note targets, PCM for Talk/telephony - Voice-note output: Google PCM is wrapped as WAV and transcoded to 48 kHz Opus with `ffmpeg` OpenClaw sends Gemini 3.8 TTS through the Interactions API (`POST /v1beta/interactions`) with `store: false`. Google also documents 3.8 on `generateContent`; Interactions is OpenClaw's routing choice, not a model requirement. OpenClaw asks for headerless 24 kHz PCM (`audio/l16`) and still wraps that PCM locally. `audioProfile` and `personaPrompt` are sent as `speech_metadata.style`, `speakerName` as the structured `speaker` label, and none of them are spoken as part of the transcript. Momentary 3.8 vocal tags use angle brackets, such as `` or ``; sustained delivery such as whispering belongs in `audioProfile`. Set `speakers` to exactly two `{ speaker, voice, style? }` entries to cast a dialogue. Only lines that start with one of those two names followed by a colon begin a turn (`Puck: Hello` and `Puck:Hello` both count), and the name is not spoken. Every other line is spoken as part of the current turn, including ordinary colon-prefixed prose such as `Budget: 10 dollars` and any unconfigured label such as `Alice: Hi`. Words before the first label are spoken by that first speaker, not dropped. A transcript with no configured labels stays on the single-voice path. Multi-speaker dialogue requires `gemini-3.8-flash-tts` or `gemini-3.8-flash-lite-tts`. ```text Puck: Headphones on. We opened it. Kore: It is waiting at the maintainer gate. ``` Gemini 3.1 and 2.5 preview TTS still use `generateContent`. Those models keep the older behavior: `audioProfile` is prepended to the transcript, and expressive tags use square brackets such as `[whispers]`. An unknown `gemini-3.8-*-tts` id fails closed instead of being sent to `generateContent`. For lowest-latency spoken conversations, use the Google realtime voice provider backed by the Gemini Live API instead of batch TTS. To use Google as the default TTS provider: ```json5 { tts: { auto: "always", provider: "google", providers: { google: { model: "gemini-3.8-flash-tts", speakerVoice: "Kore", audioProfile: "Speak professionally with a calm tone.", speakers: [ { speaker: "Puck", voice: "Puck", style: "bright" }, { speaker: "Kore", voice: "Kore", style: "whispered" }, ], }, }, }, } ``` Gemini API TTS uses natural-language prompting for style control. Set `audioProfile` for a reusable delivery style. On Gemini 3.8 that style is `speech_metadata.style` and is not read aloud. On Gemini 3.1 and 2.5 preview models it is still prepended to the spoken text. Set `speakerName` when the performance needs a named speaker; 3.8 sends it as the structured `speech_metadata.speaker` label, never as style text, and the selected voice stays the single configured voice. Gemini 3.1 and 2.5 preview TTS accept expressive square-bracket audio tags in the text, such as `[whispers]` or `[laughs]`. Gemini 3.8 uses angle-bracket vocal tags such as `` instead. To keep tags out of the visible chat reply while sending them to TTS, put them inside a `[[tts:text]]...[[/tts:text]]` block: ```text Here is the clean reply text. [[tts:text]] Here is the spoken version. Enjoy.[[/tts:text]] ``` A Google Cloud Console API key restricted to the Gemini API is valid for this provider. This is not the separate Cloud Text-to-Speech API path. ## Realtime voice The bundled `google` plugin registers a realtime voice provider backed by the Gemini Live API for backend audio bridges such as Voice Call and Google Meet. Talk and Discord expose Google's prebuilt voices in their voice catalogs. During an active Talk or Discord call, use `talk_voice` to select a new voice. OpenClaw reconnects with that voice while preserving the conversation and unfinished agent work; saved voice defaults stay unchanged. See [Discord voice changes](/channels/discord/voice-follow). | Setting | Config path | Default | | --------------------- | ------------------------------------------------------------------- | ------------------------------------------------------------------------------------- | | Model | `plugins.entries.voice-call.config.realtime.providers.google.model` | `gemini-3.1-flash-live-preview` | | Voice | `...google.voice` | `Kore` | | Temperature | `...google.temperature` | (unset) | | VAD start sensitivity | `...google.startSensitivity` | (unset) | | VAD end sensitivity | `...google.endSensitivity` | (unset) | | Silence duration | `...google.silenceDurationMs` | (unset) | | Activity handling | `...google.activityHandling` | Google default, `start-of-activity-interrupts` | | Turn coverage | `...google.turnCoverage` | Google default, `audio-activity-and-all-video` | | Disable auto VAD | `...google.automaticActivityDetectionDisabled` | `false` | | Session resumption | `...google.sessionResumption` | `true` | | Context compression | `...google.contextWindowCompression` | `true` | | API key | `...google.apiKey` | Falls back to `models.providers.google.apiKey`, `GEMINI_API_KEY`, or `GOOGLE_API_KEY` | Example Voice Call realtime config: ```json5 { plugins: { entries: { "voice-call": { enabled: true, config: { realtime: { enabled: true, provider: "google", providers: { google: { model: "gemini-3.1-flash-live-preview", speakerVoice: "Kore", activityHandling: "start-of-activity-interrupts", turnCoverage: "audio-activity-and-all-video", }, }, }, }, }, }, }, } ``` Google Live API uses bidirectional audio and function calling over a WebSocket. OpenClaw adapts telephony/Meet bridge audio to Gemini's PCM Live API stream and keeps tool calls on the shared realtime voice contract. Leave `temperature` unset unless you need sampling changes; OpenClaw omits non-positive values because Google Live can return transcripts without audio for `temperature: 0`. Gemini API transcription is enabled without `languageCodes`; the current Google SDK rejects language-code hints on this API path. Gemini 3.1 Live accepts conversational text through realtime input and uses sequential function calling. OpenClaw omits the older `NON_BLOCKING`, function response scheduling, and affective-dialog fields for this model. Prefer `thinkingLevel`; configured positive `thinkingBudget` values are mapped to the nearest supported level, while `-1` leaves Google's default in place. See the [Gemini Live capability comparison](https://ai.google.dev/gemini-api/docs/live-api/capabilities). Gemini 3.8 Live (`gemini-3.8-live`) keeps the async function-calling contract and rejects any thinking config, so OpenClaw sends none for it. Gemini 3.8 Live Extended Thinking (`gemini-3.8-live-extended-thinking`) requires `NON_BLOCKING` tools, rejects function response scheduling, and abandons a call after an interim response, so OpenClaw sends one final result per agent consult without a "working" interim. Configure its reasoning depth with `thinkingLevel` (`low`, `medium`, or `high`; `minimal` maps to `low`), or a positive `thinkingBudget` mapped to the nearest level. Spoken filler has its own utterance boundary while the interaction remains in progress; OpenClaw finalizes that transcript and audio but keeps the response active until Google reports the interaction as idle. On this model an explicit stop or barge-in interrupts generation through a short client-content turn that tells the model it was interrupted (an empty turn makes it resume). Cancelling the current generation is reliable, but the silence that follows is best effort: the model may still resume or start another response, so treat a stop as "stop this reply", not a guarantee of silence. Other Gemini Live models interrupt only through server-side voice activity detection. See the [Gemini 3.8 Live thinking guide](https://ai.google.dev/gemini-api/docs/live-api/thinking). Control UI Talk supports Google Live browser sessions with constrained one-use tokens. In Video Talk, the browser sends bounded JPEG frames directly to Google Live at the provider's maximum of one frame per second. The `describe_view` function reports whether that camera stream is active. Camera frames do not pass through the Gateway. Backend-only realtime voice providers can also run through the generic Gateway relay transport, which keeps provider credentials on the Gateway. For maintainer live verification, run `OPENAI_API_KEY=... GEMINI_API_KEY=... node --import tsx scripts/dev/realtime-talk-live-smoke.ts`. The smoke also covers OpenAI backend/WebRTC paths; the Google leg mints the same constrained Live API token shape used by Control UI Talk, opens the browser WebSocket endpoint, sends the initial setup payload plus a JPEG frame, and verifies a text response and `describe_view` function roundtrip. The OpenAI path also performs a synthesized PCM24 speech-to-response audio roundtrip; pass `--openai-audio-cycles 3` for a short repeated lifecycle soak. ## Advanced configuration Gemini Interactions is an opt-in alternative to the default `google-generative-ai` transport. Register a provider with `api: "google-interactions"` and select models through that provider ID: ```json5 { models: { mode: "merge", providers: { "google-interactions": { baseUrl: "https://generativelanguage.googleapis.com/v1beta", apiKey: "***", api: "google-interactions", models: [ { id: "gemini-3.8-flash", name: "Gemini 3.8 Flash (Interactions)", reasoning: true, input: ["text", "image"], contextWindow: 1048576, maxTokens: 65536, }, ], }, }, }, agents: { defaults: { model: { primary: "google-interactions/gemini-3.8-flash" } }, }, } ``` This transport is stateless: OpenClaw sends `store: false`, does not retain a server-side interaction ID, and replays the needed conversation context on each request. Explicit Gemini `cachedContent` handles are not supported on this route; use `google-generative-ai` for that feature. Interrupted text replies use the normal transient-error retry and failover policy. Malformed completed tool-call arguments remain rejected. For direct Gemini API runs (`api: "google-generative-ai"`), OpenClaw passes a configured `cachedContent` handle through to Gemini requests. - Configure per-model or global params with either `cachedContent` or legacy `cached_content` - Params from a more specific scope (model-level over global) always win. Within the same scope, if both keys are set, `cached_content` wins. Use only one key per scope to avoid surprises. - Example value: `cachedContents/prebuilt-context` - Gemini cache-hit usage is normalized into OpenClaw `cacheRead` from upstream `cachedContentTokenCount` ```json5 { agents: { defaults: { models: { "google/gemini-2.5-pro": { params: { cachedContent: "cachedContents/prebuilt-context", }, }, }, }, }, } ``` The optional `google-gemini-cli` runtime uses Gemini CLI `stream-json` output by default and normalizes usage from the final `stats` payload. Legacy `--output-format json` overrides still use the JSON parser. - Streamed reply text comes from assistant `message` events. - For legacy JSON output, reply text comes from the CLI JSON `response` field. - Usage falls back to `stats` when the CLI leaves `usage` empty. - `stats.cached` is normalized into OpenClaw `cacheRead`. - If `stats.input` is missing, OpenClaw derives input tokens from `stats.input_tokens - stats.cached`. If the Gateway runs as a daemon (launchd/systemd), make sure `GEMINI_API_KEY` is available to that process (for example, in `~/.openclaw/.env` or via `env.shellEnv`). ## Related Choosing providers, model refs, and failover behavior. Shared image tool parameters and provider selection. Shared video tool parameters and provider selection. Shared music tool parameters and provider selection. Gemini CLI backend setup and runtime details. Audio bridge that consumes the Gemini Live realtime voice provider.