Image: GitHub
Why it mattersA team that already shares a repo can now share the agents working in it. The work runs on each person's machine, not in one shared sandbox.
Most coding-agent apps run one agent on your own computer. Beluga, a new open-source desktop app from Beluga Labs, runs several at once. Each one gets its own folder and branch. The app can also hand a task to a friend's agent on the friend's own Mac.
The repository went public on 2026-10-16. It has 554 stars and 53 forks on GitHub. The code is TypeScript under the MIT license. Listed platforms are macOS 13+ on Apple Silicon, Linux x64 and Windows 10+ x64.
What the app does
The main idea is one task, one worktree. Every task gets its own git worktree, its own branch and its own agent. They all run at the same time. The user watches each one in the main window. Steps, sub-agents and screenshots stream live. The user can step in mid-run and change the direction.
Answers come back in the same window. The README lists live tables inside the conversation. It also lists line, area, bar, pie, scatter, Gantt, heatmap and calendar charts. Diagrams and task boards show there too.
The part that is new
The feature the project leads with is peer-to-peer collaboration between agents. A user tags a friend. The friend's Beluga asks permission once. After that, the user's agent can ask the friend's agent to do work on the friend's Mac. The traffic runs end-to-end encrypted between the two machines. Beluga Labs does not stand in the middle.
A separate feature called Sentinel runs work on a schedule, or when an event fires. It leaves a report when the work finishes. The user can also put the whole app on a Linux or macOS server. Then they reach it from a phone or a browser.
What agents it can drive
The main agent is Anthropic's Claude Code. The user signs in with their own subscription. The README also lists local models through Ollama. Anthropic-compatible endpoints work too. OpenCode is supported with a custom model setup. GitHub Copilot on the free tier is listed. A Codex plug-in is marked as coming soon.
What the README does not say
The project reports no usage numbers. There are no named customers. There is no benchmark against Cursor or Claude Code's own cloud sessions. There is no pricing. The agent subscription is the user's own. So the usage limits are the ones Anthropic or GitHub already set on the user's account. Sign-off on a task is manual. The README does not describe an automatic check that catches a wrong edit before it lands in the shared worktree.
Why this matters for a working developer
A small team that already shares a repository can now share the agents working in it. The work does not go through a vendor's cloud. Each agent uses the owner's own subscription. So one person's token spend does not land on another. The worktree-per-task layout keeps agents out of each other's way. A long-running job on one branch does not stop another agent from touching the main one.
The limit is on who the team trusts. A friend's agent reaches the friend's local files once the user approves the connection. The review of what that agent is about to do is the user's job, not the tool's. For a solo developer the P2P part is unused. The pitch then narrows to parallel worktrees and live charts in the chat. Cursor and Claude Code's cloud sessions already offer those in different shapes.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
