Gemini Provider

Rensei integrates with Google Gemini models through the gemini provider. The same catalog entry supports every auth mode - the resolver picks the execution path based on the auth mode you configure:

There is one provider in the catalog - gemini - not two. Selecting a key-based auth mode gives you cloud-capable API-direct execution; selecting local or host-session runs the model through the Antigravity agy CLI on your local daemon host, under your own Google subscription (no API key required). At dispatch the resolver rewrites a local/host-session Gemini profile to the agy-cli runtime provider.

Provider Summary

Available Models

Rensei's catalog includes:

• Gemini 3.5 Flash (gemini-3.5-flash) - Flagship agentic/coding model; 1M context, thinking support.
• Gemini 3.1 Pro Preview (gemini-3.1-pro-preview) - Deep reasoning (thinking always on), preview quality.
• Gemini 3.1 Flash Lite (gemini-3.1-flash-lite) - Cheapest sub-agent tier; low latency.
• Gemini 2.5 Pro (gemini-2.5-pro) - 2M context, thinkingBudget support, multimodal.
• Gemini 2.5 Flash (gemini-2.5-flash) - Speed/cost trade-off with optional thinkingBudget.
• Gemini 2.5 Flash Lite (gemini-2.5-flash-lite) - Standard-effort triage / sub-agent tier.

Run rensei catalog list --provider gemini to see current models.

Key-based auth modes (API-direct, cloud dispatch)

These modes use the platform's in-process Go HTTP client to call the Gemini API directly. They support cloud sandbox dispatch.

BYOK

Bring your own Google Cloud API key:

1. Get an API key from Google AI Studio.
2. In Settings → Integrations, click Add Provider and select Google Gemini.
3. Paste your API key and click Test Connection.
4. In Settings → Model Profiles, create a profile with auth mode byok and select a Gemini model.

Your org is billed directly by Google Cloud.

Metered

Platform supplies a managed API key. Your org is billed through Rensei's cost ledger. Metered mode must be enabled for your organization by the platform operator - see Auth Modes for details.

1. In Settings → Model Profiles, create a profile with auth mode metered and a Gemini model.
2. Dispatch runs if your org is entitled; contact Rensei support to enable metered mode.

Shared

Free-tier platform key for trials. No setup required; quota is soft-limited per org per day.

1. In Settings → Model Profiles, create a profile with auth mode shared and a Gemini model.
2. Dispatch runs if quota allows; returns SHARED_QUOTA_EXCEEDED otherwise.

Local / host-session auth mode (runs via your Antigravity agy CLI)

When you select auth mode local (or host-session), the resolver rewrites the runtime provider to agy-cli and donmai launches the Antigravity agy binary installed on your host - under a pseudo-terminal - instead of calling the Gemini API in-process. The CLI runs under your own OAuth subscription: no API key is injected by the platform, and your usage bills to your Google account, not Rensei.

Capability comparison: agy path vs API-direct path

Choose local/host-session to run Gemini under your own subscription with no platform key. Choose byok/metered/shared for cloud dispatch, per-profile model selection, and platform cost accounting.

Prerequisites for the agy path

Complete all of the following on the host machine that runs the donmai daemon:

agy --version

agy   # complete the OAuth login flow, then quit

rensei daemon start
rensei daemon status
rensei host provider list   # confirm `agy-cli` shows `ready`

Create a local-mode profile

Via the UI: Settings → Model Profiles → New Profile → Provider Google Gemini, Auth mode local, Model e.g. gemini-2.5-pro.

Via CLI:

rensei profile create \
  --name "gemini-local" \
  --provider gemini \
  --model-id "gemini-2.5-pro" \
  --auth-mode local \
  --scope project \
  --project-id my-project

Capability matrix (agy path)

The agy path streams the model's plain-text output as assistant events and recovers tool-use/tool-result structure best-effort from agy's on-disk transcript. The final result is captured via a structured-output instruction the provider appends to the prompt.

Credential flow (agy path)

agy OAuth session (you ran `agy login` on the host)
  └── ~/.gemini/  ──►  agy reads its own session per invocation
        (Rensei injects NO key - host-session is fail-open)

The platform does not inject any credential into agy. This mirrors the host-session pattern the claude/codex providers use - the host supplies the credential (here, an OAuth session); the platform supplies session routing.

Configuration (API-direct path)

Gemini supports safety settings and other Google-specific options via the gemini config namespace. These apply to the API-direct execution path:

{
  "providerConfig": {
    "gemini": {
      "safetySettings": [
        {
          "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
          "threshold": "BLOCK_NONE"
        },
        {
          "category": "HARM_CATEGORY_HARASSMENT",
          "threshold": "BLOCK_MEDIUM_AND_ABOVE"
        }
      ],
      "temperature": 0.7,
      "topK": 40,
      "topP": 0.95
    }
  }
}

Safety settings

Gemini applies safety filters by default. Override to suit your use case:

• BLOCK_NONE - No filtering.
• BLOCK_LOW_AND_ABOVE - Block low and above (strictest).
• BLOCK_MEDIUM_AND_ABOVE - Block medium and above (default).
• BLOCK_HIGH_AND_ABOVE - Block only high-severity (lenient).

Multimodal input (Vision)

Gemini supports image and video inputs. If your workflow includes images, Rensei automatically encodes them and sends them to Gemini's vision API. Example: analyze a screenshot in a workflow:

nodeId: analyze_screenshot
nodeType: action/llm.inference
config:
  profileId: prof_gemini_vision
  systemPrompt: "You are an expert UI/UX analyst. Analyze this screenshot."
  inputs:
    - type: image
      source: $steps.capture_screenshot.output.image_url

Context window

Gemini 2.5 and 3.x models support large context windows (1M-2M tokens). Override the default window:

{
  "providerConfig": {
    "gemini": {
      "contextWindow": 500000
    }
  }
}

Example profiles

API-direct (BYOK, cloud dispatch)

rensei profile create \
  --name "gemini-vision" \
  --provider gemini \
  --model-id "gemini-2.5-flash" \
  --auth-mode byok \
  --scope project \
  --project-id my-project

Local (agy, your subscription)

rensei profile create \
  --name "gemini-local" \
  --provider gemini \
  --model-id "gemini-2.5-pro" \
  --auth-mode local \
  --scope project \
  --project-id my-project

Then route work to the profile:

rensei routing set \
  --scope project \
  --project-id my-project \
  --work-type development \
  --profile-id prof_gemini_local

Pricing

View pricing in the model catalog:

rensei catalog show gemini-2.5-flash --format json | jq .pricing

For metered mode, cost events emit automatically with the org ID for billing. The agy path bills to your own Google subscription and reports no platform cost events.

Troubleshooting

"No metered key is configured for provider 'gemini'"

The platform's metered key pool for Gemini is not configured. Contact your Rensei operator - deployment-level key pool setup is covered in the operator docs.

"Invalid API key"

Your BYOK key is expired or malformed. Generate a new key from Google AI Studio and update Settings → Integrations.

"Organisation is not entitled to use metered auth mode"

The org's metered_enabled feature flag is not set. Contact Rensei support to enable metered mode for your organization.

AUTH_MODE_REQUIRES_LOCAL_CAPACITY (agy path)

You are dispatching a local/host-session profile to a cloud sandbox. Either dispatch from a machine with the donmai daemon running, or switch to a key-based auth mode (byok/metered/shared) for cloud dispatch.

ErrProviderUnavailable: agy CLI "agy" not on PATH (agy path)

The agy binary is not installed or not on $PATH when the daemon starts. Install Antigravity, confirm agy --version, then restart the daemon:

rensei daemon restart
rensei host provider list   # `agy-cli` should show `ready`

Sessions dispatch but produce no output / activities (agy path)

agy is probably not logged in. Run agy interactively once on the host to complete the OAuth login, then rensei daemon restart. (The agy path uses no GEMINI_API_KEY - it relies on your OAuth session.)

Vision input not processed (API-direct mode)

Ensure the image URL is accessible and in a supported format (JPEG, PNG, GIF, WebP).

Safety filter blocking requests (API-direct mode)

Review your safetySettings config. For development/testing, use BLOCK_NONE; for production, tune thresholds to your use case.

Further Reading

• Auth Modes - Full auth mode reference, including local and host-session
• Local Sandbox - Local pool setup and daemon lifecycle
• Model Catalog & Routing - Profile and routing management
• Antigravity - The agy CLI (successor to the EOL Gemini CLI)
• Google AI Studio - Get a GEMINI_API_KEY for the API-direct path
• Google Gemini Docs - Official API reference