Beluga IDE: AI-Native Mac App with Peer-to-Peer Agent Communication

What changed

Beluga has released version v2.3.2-mac, an AI-native IDE for macOS that emphasizes peer-to-peer connections and is designed specifically for AI development workflows. The application allows users to work on multiple tasks concurrently, with each task assigned its own dedicated worktree, branch, and agent. These agents operate side-by-side, providing live updates on their progress, including steps, sub-agents, and screenshots, which can be steered mid-run.

Key features include enhanced visualization of agent work, with live tables, charts, diagrams, and task cards displayed directly within the conversation interface. A significant aspect is the peer-to-peer agent communication, enabling an agent to request tasks from a friend's agent on their Mac, requiring explicit user approval. This functionality is facilitated by direct computer-to-computer connections, end-to-end encrypted, without relying on a central server.

The 'Sentinel' feature allows users to define recurring or event-driven tasks that run on a schedule or in response to specific triggers, generating reports upon completion. Beluga supports installations on macOS 13+, Linux x64, and Windows 10+ x64, with self-updating capabilities. Installation involves downloading the application, with specific instructions for each operating system, including workarounds for macOS and Windows security prompts. Integration with Claude Code is also highlighted, requiring Git for Windows on Windows installations.

For collaborative work, users can enable the 'Friends' feature, turn it on, and add friends using a provided code. The system allows agents to interact with friends' agents for tasks that require access to their local environment, such as running tests or accessing specific files, with clear permission prompts for the friend.

Why it matters for builders

Beluga introduces a paradigm shift for AI developers by embedding AI agents directly into the IDE, transforming how projects are managed and executed. The ability to assign each task its own isolated worktree and agent simplifies complex project structures and reduces interference between concurrent operations. The live, visual feedback on agent progress and the interactive data visualizations (tables, charts, diagrams) provide unprecedented insight into AI agent operations, making debugging and steering more intuitive.

Furthermore, the peer-to-peer collaboration feature directly addresses the challenges of distributed AI development and task delegation. Builders can leverage their colleagues' local environments and resources securely, accelerating development cycles and enabling specialized task execution without centralizing sensitive data or credentials. The Sentinel feature automates background tasks and monitoring, freeing up developers to focus on core development.

Practical impact

Developers can now download and install Beluga on their macOS, Windows, or Linux machines. For Mac users, this involves downloading the .dmg file and following instructions to bypass macOS security prompts if necessary. Windows users will run an installer and may need to install Git for Windows. Linux users can use AppImage, .deb, or .rpm packages.

To begin, users can sign in to Claude Code and start describing tasks, which Beluga will then manage within dedicated worktrees. For collaboration, enabling the 'Friends' feature and adding contacts will allow for peer-to-peer agent interactions. Builders can experiment with the Sentinel feature by defining simple automated tasks, such as monitoring a local service or running scheduled scripts.

The ability to run Beluga on a server (Linux or macOS) and access it via a phone app or browser opens up possibilities for continuous integration and remote development workflows. The beluga setup command on a server simplifies initial configuration, including signing into Claude, naming the machine, and pairing mobile devices.

Caveats and source limits

The source material indicates that Beluga is a fresh release and not yet notarized by Apple, which requires users to manually bypass security settings on macOS. The Windows installer is also not signed, necessitating a similar manual override. The documentation mentions that Git for Windows is a prerequisite for Claude Code on Windows. While the application supports peer-to-peer connections, the specifics of network configuration and potential firewall issues are not detailed. The source does not provide information on pricing, specific performance benchmarks, or the exact range of AI models supported beyond its integration with Claude Code. The 'Sentinel' feature's full capabilities and limitations are also not exhaustively detailed.