Give every agent in Herdr its own Vercel Sandbox

Terminal coding agents such as Claude Code, Codex, and OpenCode can now operate within separate Vercel Sandboxes, managed through Herdr, a tmux-like interface. The plugin uploads project files to a new Sandbox, installs the selected agent, and opens a connected pane for interaction, all without affecting the local machine. This allows multiple agents to run simultaneously without performance degradation. Each agent operates in its own isolated environment, ensuring no cross-contamination of processes or files.

The plugin provides machine-readable outputs for every action, returning results in JSON format. This enables scripting or automation of tasks such as starting agents, monitoring their state, or applying changes. Users can script complex workflows or delegate tasks to other agents, leveraging the structured data returned by the plugin. The integration with Herdr’s pane system ensures a familiar terminal-like experience while maintaining remote execution.

Several safeguards are implemented to protect user data and system integrity. The plugin displays all files slated for upload before the first operation, allowing users to review or modify the selection. Remote changes are returned as Git patches, which users must explicitly apply. A manual confirmation—typing DELETE—is required to permanently remove a Sandbox, preventing accidental data loss. These measures ensure transparency and control over agent operations.

The feature supports verified versions of Claude Code (2.1.220), Codex (0.146.0), and OpenCode (1.18.9) by default. To use it, developers must run `vercel link` in their project repository and initiate an agent from Herdr’s action menu. The process involves two steps: a dry run to preview changes, followed by a second invocation to execute the operation. Changes made by agents can be pulled back to the local machine as Git patches. Usage is billed at standard Vercel Sandbox rates, with documentation available for further details.