ChatterKey v4.6.0 — Gemini, directly
ChatterKey v4.6.0 — Gemini, directly
macOS 14+ · Apple Silicon download · Build 10
Use your Gemini API key directly in ChatterKey. This release adds Google Direct alongside OpenRouter while keeping the same single-model, single-request dictation and voice-edit architecture.
What's new since v4.5.0
- Google Direct: audio goes straight to Google's OpenAI-compatible endpoint using your Gemini API key. An OpenRouter account is not required for this connection.
- Connection picker: choose Google Direct or OpenRouter in Settings and onboarding. New installations default to Google Direct with
gemini-3.5-flash-lite. - Separate credentials and preferences: each connection keeps its own Keychain account, model ID, and cost rates. Switching connections never copies an API key to the other provider.
- Provider-specific requests: Google uses native model IDs and
reasoning_effort; OpenRouter keeps its own model IDs and routing parameters. Google requests omit OpenRouter-only headers and fields. - Safe upgrades: existing OpenRouter setups retain their selected connection, model, and rates. Choosing Google Direct is explicit. Local vocabulary, snippets, writing preferences, and history remain available.
- Updated documentation: setup instructions, privacy/security disclosures, README, website, and download links now cover both connections.
What stays the same
- One audio-capable model request per attempt for all eight writing modes and Magic Voice Edit.
- Audio, instructions, and selected text for voice edits travel together in that request.
- No separate transcription/polishing stage, automatic repair, automatic retry, or automatic connection fallback. Retry remains a user action.
- Endpoint pinning and redirect rejection protect provider credentials.
- Optional live preview stays on-device and does not add a cloud request.
Set up Google Direct
- Quit the previous app, extract
ChatterKey-v4.6.0.zip, and replace the app in Applications. - Open Settings → AI Provider → Connection → Google Direct.
- Enter your own Gemini API key in the Gemini API key field. Do not paste an OpenRouter key here.
- Keep
gemini-3.5-flash-liteas the model, then Save. - Test Connection checks API access. Dictate a short clip to check the selected model's audio access.
If you prefer OpenRouter, keep or select that connection and use its own key. Model IDs and cost rates remain independently editable. Dashboard costs are estimates, not provider invoices.
Validation
- The existing Swift 6 regression harness covers both connections, all writing modes with cleanup on/off, audio payloads, one-request voice edits, connection tests, migration, model/rate isolation, invalid responses, cancellation, authentication/quota errors, and cost estimates.
- Debug/release builds, public-source scans, app-signature checks, website checks, and release-archive integrity checks are used for this release.
- Request counts and failure handling are verified with mock transports. No new live Google/OpenRouter audio test or clean-account macOS test was performed as part of release preparation; measured provider latency and accuracy are not claimed.
Download and distribution limits
- Requirements: Apple Silicon Mac, macOS 14+, your own API key, and Microphone/Accessibility permissions. Speech Recognition permission is optional for live preview.
- Binary:
arm64, not a universal Intel/Apple Silicon build. - Signing: ad-hoc signed community-test build, not Developer ID signed or Apple-notarized. Gatekeeper may block it. Do not disable system-wide security protections to install it.
- Checksum:
ChatterKey-v4.6.0.zip.sha256accompanies the ZIP. It verifies download integrity, not Apple approval or publisher identity.
After downloading both files to the same directory:
shasum -a 256 -c ChatterKey-v4.6.0.zip.sha256Short announcement
ChatterKey v4.6.0: Gemini, directly. Bring your Gemini API key and use Google Direct without an OpenRouter account. OpenRouter remains optional, keys and model/rate preferences stay separate, and every dictation or voice-edit attempt still uses one model request. Available as an Apple Silicon community-test build; not Apple-notarized.