wickrunAI: Desktop App for Orchestrating Multiple AI Models

What changed

wickrunAI has been released as an open-source desktop application for Windows, macOS, and Linux, built with Electron and React, and featuring a Capacitor-based Android client. The core functionality revolves around orchestrating various AI models, including open-source, closed-source, paid, and free services, under a unified interface. A key feature is the "bring your own key" (BYOK) system, where API keys are encrypted and stored locally using the operating system's secure keystore (Windows DPAPI, macOS Keychain, Linux libsecret), never leaving the user's machine. This ensures enhanced privacy and security for API credentials.

The application introduces an automatic task handover mechanism. If an AI model fails mid-task due to issues like quota limits or errors, wickrunAI can seamlessly transfer the ongoing task to the next model in a pre-defined failover list. This process preserves the saved progress, allowing the task to resume from where it left off. The failover order is user-defined, enabling granular control over which models are used and in what sequence. The system supports session, project, and app-level scoping for failover lists, allowing for inheritance or overriding of configurations.

Recent updates include improvements to desktop interactions, such as handling IME composition keys and differentiating between Chat/Work switching and Stop controls. Streaming behavior has been refined to follow new content until scrolling or text selection, with a "Jump to latest" option to resume following. The application can now run in the background when the window is closed, accessible via the system tray or macOS Dock, with tasks continuing as long as the computer remains awake. Existing task recovery mechanisms handle app exits and restarts.

Why it matters for builders

For AI builders, wickrunAI provides a consolidated platform to experiment with and deploy a wide array of AI models without the complexity of managing multiple API integrations and credentials. The BYOK approach is particularly beneficial for developers who prioritize data privacy and security, as sensitive API keys are handled locally. The automatic failover capability is a significant advancement for building resilient AI agents and workflows, ensuring that tasks can continue uninterrupted even when individual model services encounter problems. This reduces the need for manual intervention and improves the overall reliability of AI-powered applications.

Practical impact

Builders can get started with wickrunAI by downloading and installing the application, then adding their API credentials through the settings. A quick setup involves adding a single API key, with options to use presets like OpenRouter for an immediate free route. For more advanced use, users can configure multiple routes in a failover list, integrate file tools via a working directory, or set up a search service. The application also supports cross-device synchronization of conversations, projects, and tasks via a user-chosen folder (e.g., Syncthing, cloud-synced directories), with data encrypted using scrypt and AES-256-GCM. Conversations can be exported in Markdown or JSON formats, with API keys and credential IDs stripped for security. The Android client allows for remote control of desktop tasks over a local network, though it has not yet been verified on physical devices.

Caveats and source limits

The source material indicates that the Android client has not yet been verified on a physical device. While the application supports various AI models, specific details on the exact number or types of models officially supported beyond general categories (open, closed, paid, free) are not provided. Pricing for API access is managed by the individual model providers, as wickrunAI itself is a BYOK client and does not handle billing. The comparison table lists features for other applications but notes that some answers are unverified ('—'), meaning those products might have the feature without explicit confirmation in this source. The source does not include independent benchmark results for wickrunAI's performance or failover efficiency.