Model configuration

View source on GitHub

Use Settings → Models & Usage → Custom Models in the standalone desktop app. Older versions call this page Models. The editor and Import action validate configuration and encrypt credentials before saving.

Model files and import

Desktop models live at ~/.gemini/antigravity/custom_models.json. Import accepts a models object, a model array, or provider groups with inherited defaults, as JSON or base64-encoded JSON. Existing internal names are updated; new names are added. Provider-group defaults become independent model entries after import.

This secret-free example targets a local OpenAI-compatible server. Replace YOUR_LOCAL_MODEL_ID with an ID returned by Discover local or your server:

{
  "models": [
    {
      "name": "models/local-example",
      "displayName": "Local example",
      "provider": "lmstudio",
      "apiFormat": "openai",
      "apiUrl": "http://127.0.0.1:1234/v1",
      "apiKey": "",
      "externalModelName": "YOUR_LOCAL_MODEL_ID",
      "enabled": true,
      "contextWindow": 32768,
      "maxOutputTokens": 4096,
      "timeout": 120000,
      "maxRetries": 2
    }
  ]
}

Choose limits supported by your actual model. The example does not install a server or download a model. A local server that requires authentication still needs its own API key.

Discover and select models

Select a provider, enter its key and endpoint, then choose Fetch provider models. Search the returned catalog by model name or ID. Select filtered selects the current matches; selections remain checked when you change the search. Add selected saves all checked models, including selections outside the current filter.

When available, discovery imports each model’s context/output limits and image and reasoning metadata. Explicit values in Advanced settings take priority. An output limit at least as large as the context is reduced to leave input room. Missing metadata keeps the normal defaults; review limits and capabilities for your chosen model. Discovery does not generate a completion or verify tool use. See providers and model catalogs for endpoints and service-specific details.

Image inputs

Choose a model and endpoint that actually support image understanding. Discovery imports image capability metadata when the provider supplies it; otherwise the addon recognizes known vision model names, including DeepSeek-VL/VL2 and Janus/Janus-Pro. These open models require a compatible serving endpoint; their names alone do not make them available on the hosted DeepSeek API. See the DeepSeek vision guide for currently supported hosted models.

In the model editor, Advanced settings → Image input → Supported explicitly enables image attachments. Unsupported disables them; Auto detect uses the model’s default detection. JSON/import users can set supportsVision: true or false. Both values take priority over name detection. This setting cannot add vision to a text-only upstream model.

The desktop proxy forwards PNG, JPEG, GIF and WebP images as native image blocks for OpenAI-compatible and Anthropic-compatible endpoints. It preserves images in user conversation history, including images beside tool results. Inline base64, base64 data URLs and local file:// images are sent as image data; HTTP(S) image URLs are passed to the selected provider and must be publicly accessible. Private URLs requiring a separate login must be supplied as inline image data or a local file instead. Incoming Google credentials are not forwarded to image URLs. Unsupported image formats and unreadable local images return a clear error instead of being sent as text. These formats accept image input in user messages.

Desktop context trimming budgets an estimated 1024 tokens per image rather than treating base64 bytes as text tokens. Actual image cost depends on the provider, model and resolution; upstream image-size and context limits still apply.

Fields and advanced settings

Most fields are exposed under Advanced settings. Fields marked JSON/import are supported in configuration but do not have a separate editor control.

Timeout/budget fields accept 1–3600000 milliseconds. Context limiting uses a rough token estimate and removes whole oldest turns while preserving supported tool exchanges. The proxy refuses a latest turn that still cannot fit. Retries and fallback stop after output starts; they do not duplicate an already streamed answer. Renaming/deleting a model updates its saved fallback references. For parallel workloads, see fallback chains for rate-limited providers for worked examples, retry timing and the separate MCP delegation behavior.

Old DeepSeek, Kimi, Fireworks, LM Studio, and llama.cpp entries without apiFormat retain their historical Anthropic format. Select the intended format explicitly when migrating a configuration.

Subagent model selection

When Antigravity exposes the invoke_subagent tool, a conversation using either a built-in or custom model can choose an enabled configured model through Subagents[].Model. Use the model’s exact saved name, such as models/local-example from the import above; the display label and upstream provider model ID are not interchangeable with this name. Configure and enable the target model first, then include its saved name when asking the agent to delegate work.

The native values inherit, flash_lite, flash, and pro remain available. inherit (also the default when omitted) follows the calling agent’s model, including a custom model selected for that subagent. Explicit native tiers keep Antigravity’s own tier selection. Unknown, ambiguous, or disabled custom model selections produce an error instead of silently choosing another model.

The proxy gives each custom-model spawn its own temporary routing marker. The child’s first request can select its model before the parent receives the tool result, and parallel calls do not share a fallback tool-call ID. Markers are removed before forwarding the child’s prompt to a provider; replayed parent history restores the original prompt and Model even after switching between Gemini, OpenAI-compatible, and Anthropic-compatible providers.

Built-in Google/Cloud Code parents use the same routing bridge. The proxy exposes custom model selection in their tool schema and rewrites their JSON or SSE tool responses before Antigravity validates the native model tier. Cloud Code response envelopes and compressed responses are handled without dropping their metadata. An invalid selection returns an explicit error; if a stream has already started, the proxy emits a terminal SSE error instead of forwarding an invalid tool call.

Routing state is held in proxy memory, with a one-hour inactivity limit and up to 4,096 entries per routing cache. Start new subagents after restarting the proxy or when a route has expired; a child using an unavailable marker reports an error. Expired markers in parent tool history are removed so the parent can continue and spawn fresh children; its original custom selection can only be restored while the corresponding routing state remains available. This feature depends on Antigravity supplying the child prompt or a stable conversation identity, and does not add native subagent tools to clients that do not expose them. Automated routing coverage uses local HTTP fixtures. Actual Windows IDE 2.5.5 requests from both built-in Gemini and custom parents did not declare invoke_subagent. Google lists native subagents for the separate Antigravity 2.0 app and CLI in its subagent documentation.

For the IDE, use the supported MCP delegation setup. The delegate_to_model tool sends an independent text task to an exact saved custom model and returns its answer to the parent. This worker has no IDE tools, filesystem access or inherited conversation; it does not open a native child window. Native invoke_subagent routing remains available in clients that actually expose that tool.

Gateway routes

Configure Remote gateway with the gateway URL, token, and dashboard URL. Test gateway checks authenticated model listing; Import model aliases adds routes using provider: "google", apiFormat: "google", and gateway: true. Imported routes retain advertised context/output limits and image/thinking capabilities. The default imported context window is 128000; unspecified image/thinking capabilities default to false. Re-import to refresh alias metadata.

The saved connection is ~/.gemini/antigravity/gateway_connection.json. Changing the gateway origin requires entering its token again. See the gateway guide for setup, network listeners, tunnels, and the gateway’s separate configuration.

Google Cloud Code account pools

Choose Google Cloud Code account pool and enter an available model ID. Add an account before using Fetch provider models; discovery uses an enabled account. The pool is separate from the application’s vendor-managed sign-in.

1. Create your own Google OAuth client with application type Desktop app and configure consent/test users for your project. Enter its client ID and optional client secret under Sign in with your OAuth client.
2. Click Sign in with Google. This explicitly opens the browser using Google’s installed-app OAuth flow, with PKCE, state validation, and a temporary loopback callback. The requested scope is cloud-platform. You can cancel; the login attempt times out after two minutes.
3. Save the model within 10 minutes. Pending credentials remain in main-process memory until saved; the renderer receives masked fields. Repeat sign-in if an unsaved entry expires.
4. Alternatively, use Import account credentials for your own authorized_user JSON, an account object, or an array. Import itself makes no network requests. Save the model to retain the credentials encrypted.
5. Use Test account to check authorization and retrieve a quota snapshot. Pool status shows current state. Quota refresh is explicit, not periodic.

Each account has a unique stable id, optional label, enabled, project, clientId, clientSecret, refreshToken, accessToken, and expiresAt (Unix milliseconds). Up to 50 accounts are supported. Keep the ID unchanged when editing masked credentials. Account project overrides googleProject when set.

googlePool.strategy accepts round-robin, least-loaded, quota, or adaptive. Adaptive selection compares two eligible candidates using measured time to response headers, in-flight load, recent failures and fresh known quota. Pool status exposes request/success/failure counts and measured latency. Measurements and quota snapshots become stale after ten minutes; unknown values remain unknown. Remaining-quota selection uses the latest fresh retrieved snapshot. Conversation affinity can retain an eligible account. maxConcurrency is per account (1–100, default 2); cooldownMs defaults to 60000 and accepts 1–3600000. Access tokens refresh as needed; failed authorization and cooldown/concurrency state can temporarily make an account unavailable.

Structured Google RetryInfo / ErrorInfo hints and Retry-After take precedence over the configured cooldown (bounded to 31 days). Cooldowns use the upstream model ID, so local effort variants share the same limit. A cooldown does not block a different upstream model using that account; authentication failures and concurrency capacity remain account-wide. A quota snapshot disables an account globally only when all returned fresh buckets explicitly report zero. One empty Claude bucket must not disable a positive Gemini bucket. No refill time is guessed.

No third-party OAuth client credentials are bundled, and no browser/IDE/gcloud credentials are automatically extracted. Successful OAuth does not grant Cloud Code entitlement, model access, or quota. Supply an eligible account/project; Google’s location and service-access decisions remain in effect.

Effort variants and recent models

Choose Effort variants on a saved model to review and create low/medium/high copies. OpenAI-compatible models receive reasoningEffort; other formats also receive thinking budgets of 1024/4096/8192 and sufficient output budgets. Confirm your provider supports these values. The source model stays unchanged, name collisions get numbered suffixes, and saved credentials are copied privately. No provider request occurs when creating variants.

The model manager orders recently successful models first within its grouping. This history is process-local and excludes generation tests. It does not change native IDE picker ordering or the stored model identities.

Direct-proxy diagnostics and generation tests

Open Diagnostics in the desktop addon or IDE browser manager to inspect up to 200 recent requests, status, elapsed time and provider-reported input/output/ cached tokens. Missing usage is shown as unknown. Totals cover this bounded in-memory window, not billing history; restarting or clearing diagnostics removes it. Usage belongs to the final attempted model; earlier failed retry/fallback attempts are excluded from token totals. Elapsed time covers the whole request. Prompt text, response text, request URLs and credentials are not retained.

Test generation (uses quota) in the model editor sends the fixed prompt Reply with OK. with a 32-token output budget, a 30-second limit, no tools, retries or fallback models. It exercises the same request translation as normal chat and reports whether generated text was received. Some reasoning models need a larger minimum output budget and can fail this small test despite being usable. Test connection still checks the provider’s catalog/authentication without requesting generation.

Credentials and data

Desktop API keys, custom-header values, saved gateway tokens, and Google access/refresh tokens and client secrets are encrypted before persistence. Electron safeStorage is preferred; otherwise AES-256-GCM uses a generated local key at ~/.gemini/antigravity/.model-credentials-key. Keep the key with private backups. Machine-encrypted secrets are not portable configuration; enter fresh credentials when moving to another machine or when decryption fails.

Saved secrets appear as ********; leaving them unchanged preserves them on the main-process side. Changing a provider or server origin requires fresh credentials. Export removes API keys, all custom headers, and Google access/refresh tokens and client secrets. It retains ordinary settings, account labels/IDs, client IDs, and project names. Do not put secrets in extraBody, URLs, descriptions, or other ordinary fields, and review exported metadata before sharing.

The optional gateway has a separate profile at ~/.gemini/antigravity/gateway. Its credential configuration does not use Electron safeStorage; its SQLite history stores prompts, responses, tools, and replay snapshots. Consult gateway data handling before sharing that profile. Desktop model Export does not export the gateway’s database.