Give every agent in Herdr its own Vercel Sandbox

What happened

Terminal coding agents such as Claude Code, Codex, and OpenCode can now each run inside their own isolated Vercel Sandbox, orchestrated from Herdr. Herdr is a tmux-style manager that places those agents side by side in panes so you can drive several of them in one session. The core claim is isolation: nothing an agent runs or edits touches your machine. Startup is explicit. When you start an agent, a plugin uploads your project files into a new Sandbox, installs the agent there, and opens a pane connected to that remote environment. The pane is the live terminal surface for that agent, not a local shell on your laptop.

The architecture is remote execution with local orchestration. Herdr sits on your side of the glass and behaves like a pane manager. Each agent process lives in a Vercel Sandbox of its own. A Sandbox is an isolated compute environment provisioned for that agent’s run, not a shared host with the rest of your machine. The plugin’s job is the handoff: copy project files into the new Sandbox, install the chosen agent inside it, then wire a Herdr pane to that Sandbox so keystrokes and output stream as if the agent were local. Isolation is therefore both filesystem and process level. Edits, package installs, shell commands, and failed experiments stay inside the Sandbox. Your host disk and host processes stay out of the agent’s reach unless you deliberately pull results back.

The technical detail

That split matters for people who already use terminal coding agents. Claude Code, Codex, and OpenCode are useful because they can edit, build, and run commands, but the same power is the risk when the agent runs on the machine that holds production credentials, personal keys, or uncommitted work. Putting each agent in its own Vercel Sandbox turns “try a risky refactor” into something you can do without gambling the host. Running several agents side by side in Herdr panes also changes the workflow shape. You can keep one agent on a branch of work, another on tests or docs, and another on a second approach, without stacking everything into one local process tree. Builders who already think in tmux sessions get a familiar layout with a stronger isolation boundary.

Why it matters for builders

The competitive and market frame is the rise of agent orchestration tools next to cloud sandboxes. Terminal agents are no longer a single CLI on one laptop; they are concurrent workers that need places to run. Vercel Sandbox supplies the isolated runtime. Herdr supplies the multi-pane manager. Together they sit between pure local agent installs and fully remote IDE products. Other stacks offer remote dev environments or containerized agent runs, but this pairing is specific: named terminal agents, a tmux-style front end, and one Sandbox per agent. For teams that already ship on Vercel or trust Vercel for ephemeral compute, the path of least resistance is clear. For everyone else, the product still answers a general need, keep agent side effects off the developer machine while preserving a real terminal UX.

Market and competitive context

Practically, the next things to watch are how the upload, install, and pane connection behave under real projects, and how you move finished work back. Large monorepos, secrets handling, and network access from the Sandbox will decide whether this is a daily driver or a specialty mode. Watch whether multi-agent sessions stay stable when several Sandboxes run at once, and whether install steps for Claude Code, Codex, and OpenCode stay reliable as those tools update. Also watch the loop that matters most after isolation: commit, review, and bring artifacts home without turning the Sandbox into a black hole for changes.

What to watch next

Risks and open questions remain even with clean isolation. Uploading project files into a Sandbox means the remote environment receives a snapshot of your tree. That is safer for the host but raises questions about what leaves the machine, how long Sandboxes live, and who can see them. Agents can still make bad edits inside the Sandbox; isolation protects your laptop, not your branch quality. Prior art includes containerized agents, remote containers, and classic tmux multi-session workflows. Herdr’s contribution is the orchestration surface that makes one pane equal one Sandbox-backed agent rather than one more local shell. If the pane connection is solid and the upload path is predictable, the model is simple enough to stick: start agent, get isolated Vercel Sandbox, work in the pane, leave the host untouched.