OppenheimerGPT

Routes a single user prompt to multiple AI providers (OpenAI, Anthropic, Google, etc.) in parallel, executing inference calls concurrently rather than sequentially. Implements a provider abstraction layer that normalizes API schemas across different LLM endpoints, handling authentication tokens, rate limiting, and response formatting differences transparently. Uses async/await patterns to fire requests to all configured models at once, reducing total wall-clock time compared to serial API calls.

Renders multiple model responses side-by-side in a split-pane UI, with synchronized scroll position across all panes so users can compare responses line-by-line. Implements a layout engine that dynamically adjusts column widths based on number of active models and screen resolution. Highlights differences between responses (via text diffing or visual markers) to surface where models diverge in reasoning or output format.

Stores and manages API keys/credentials for multiple AI providers (OpenAI, Anthropic, Google, etc.) in a centralized credential vault, likely using macOS Keychain for encrypted storage. Implements a provider registry that maps credentials to specific model endpoints and handles token refresh/rotation for OAuth-based providers. Abstracts credential lookup so users configure once and the app automatically injects the correct token into each provider's API call.

Provides a settings interface where users enable/disable specific AI models and configure provider-specific parameters (temperature, max tokens, system prompts, etc.). Maintains a model registry that lists all supported providers and their available models, with UI controls to toggle which models are active for the current session. Stores configuration state locally (likely in a JSON or plist file) and applies settings to all subsequent inference calls.

Maintains a local history of all prompts and responses from the current session (and optionally previous sessions), allowing users to revisit past queries and model outputs. Implements a session abstraction that groups related prompts/responses together, with UI controls to browse history, search past queries, and optionally export sessions. Likely stores history in a local database (SQLite or similar) with metadata (timestamp, models used, response times).

Automatically measures and displays latency metrics for each model's response (time-to-first-token, total response time, tokens-per-second), enabling users to benchmark model performance. Collects timing data at the API call level (request sent → response received) and optionally at the token level if streaming is supported. Displays metrics in the UI alongside responses, likely with visual indicators (progress bars, timing badges) to make performance differences obvious.

Supports streaming responses from models that offer token-by-token output, rendering tokens incrementally as they arrive rather than waiting for the full response. Implements a streaming parser that handles provider-specific streaming formats (OpenAI's Server-Sent Events, Anthropic's streaming protocol, etc.) and updates the UI in real-time. Maintains separate streaming state for each model, allowing users to see responses arrive at different speeds simultaneously.

Provides UI controls to copy individual model responses to clipboard, or export multiple responses (from a single prompt across all models, or from an entire session) to file formats like Markdown, JSON, or plain text. Implements formatting logic that preserves response structure (code blocks, lists, etc.) when exporting. Supports batch export of entire sessions with metadata (timestamps, model names, parameters used).

Implements keyboard shortcuts for common actions (send prompt, switch models, copy response, export session, etc.), with customizable hotkey bindings. Likely supports global hotkeys that work even when the app is not in focus, allowing users to trigger model comparisons from anywhere on macOS. Implements a hotkey registry that maps key combinations to actions and persists custom bindings in user preferences.