1. Sakana AI introduces KAME tandem speech-to-speech architecture
Sakana AI has introduced KAME, a tandem speech-to-speech (S2S) architecture that injects real-time LLM knowledge into conversational AI. The system pairs a low-latency S2S model for immediate responsiveness with an asynchronous backend LLM that provides refined knowledge signals. This design allows the agent to speak while thinking. It bridges the gap between fast S2S models and deep-knowledge cascaded systems without adding latency penalties.
2. Developer releases DeepClaude proxy for Claude Code
A developer has released DeepClaude, an open-source proxy that integrates DeepSeek V4 Pro with Anthropic's Claude Code CLI. The tool swaps the underlying model while preserving the terminal UX, file editing, bash execution, and autonomous multi-step coding loops. By routing model API calls to DeepSeek or OpenRouter while maintaining the Anthropic WebSocket bridge, developers can run the agent loop at a significantly lower cost per million tokens.
3. Flue launches TypeScript framework for autonomous AI agents
Flue has released a new TypeScript framework designed for building programmable, autonomous AI agents. The framework uses a sandbox architecture that pairs an agent harness with a secure, isolated container workspace. This allows agents to safely edit files, execute code, spawn subagents, and run terminal commands. Developers can deploy the framework via CLI or HTTP to build custom workflows without relying on generic hosted systems.
4. Nexu releases Open Design open-source design workflow
Nexu has released Open Design, an open-source, local-first alternative to Anthropic's Claude Design. The tool wires existing local coding agents, such as Claude Code, Codex, or Cursor Agent, into a skill-driven design workflow. It uses a daemon to build a real on-disk project folder with seed templates and layout libraries, enforcing pre-flight checks and self-critique before rendering artifacts in a sandboxed iframe. The web layer can be deployed to Vercel, allowing developers to bring their own keys at every layer.
5. Developer releases Agent-desktop automation CLI
A developer has released Agent-desktop, a native desktop automation CLI for AI agents. Instead of relying on token-heavy and fragile screenshot scraping, the tool uses structured OS accessibility APIs like macOS Accessibility, Windows UI Automation, and Linux AT-SPI. This provides agents with structured UI information, similar to how Playwright operates on the web. The approach enables more reliable and efficient desktop automation for computer-use agents.
6. Developer ports Apple's SHARP 3D model to the browser
A developer has released a browser-based implementation of Apple's SHARP single-image 3D Gaussian splatting model. The project exports the predictor to ONNX and runs it entirely client-side using onnxruntime-web with the WebGPU execution provider. Users can drop in an image and generate a downloadable .ply file locally without sending data to a server. The implementation requires a large initial cache load but completes inference in seconds on recent hardware.