VS Code in the Browser, with Claude Code, Codex and opencode Ready

Coding agents are the best thing to happen to a backlog in years — and running them on a laptop is the worst part of using them. You install a CLI, export an API key, and hand your working tree to a process that edits files while you are trying to use the machine. Long runs chew the battery. Two runs on one checkout trample each other. And when the agent finishes at 2am, its diff is just sitting there in your uncommitted changes, unexplained.

There is a simpler shape for this: run the agent somewhere disposable, and keep a real editor open beside it. That is what an Oyren sandbox is — coding agents on our servers, in a disposable sandbox with your repo checked out. Use any supported CLI — Claude Code, Codex, opencode or Cursor — and open a real VS Code in the browser beside it.

From a repo to a draft PR

The workflow is deliberately short:

1. Connect a GitHub repo. Private or public — the sandbox clones it for you.
2. Pick a harness and a model. The CLI is one choice; the model behind it is a separate one. Bring your own key or use credits.
3. Describe the task. Oyren boots an isolated container, starts the agent, and gets out of the way.
4. Review the draft PR. The agent commits as it goes, so nothing lives only inside a container, and it opens a draft pull request for you to read.

The part that makes this feel like your desk rather than a job queue is the editor. A real VS Code runs in the browser, on the same machine as the agent — same filesystem, same session. If the agent takes a wrong turn, you open the file and fix it by hand. Nothing is lost in the handoff between "the agent's machine" and "yours", because they are the same machine.

That is also what separates this from watching a transcript scroll by. You can run the tests yourself, poke the dev server, and read the diff in an editor instead of a log.

Pick your harness

If you want to run Claude Code in the cloud, pick it and it is there — no install, no login dance on a remote box. The same goes for OpenAI's Codex, the open-source opencode, and Cursor's headless agent. The roster is deliberately model-agnostic: opencode in particular runs whatever model you point it at, which is how Kimi K3 wrote a post on this very blog.

What keeps it safe

A sandbox that can edit your repo needs rules, and these are on the record:

• Short-lived, repo-scoped tokens. Agent sandboxes get roughly an hour of access to one repository. Nothing else.
• Every step on the record, with its cost. You can see each command the agent ran and what it spent.
• Stops and asks when it is not sure. Ambiguous or destructive steps wait for you.
• Credentials encrypted at rest with AES-256-GCM — and your work is never used for training, on any plan, including free.

A GitHub Codespaces alternative? Yes and no

Codespaces answers "give me a dev machine in the cloud", bills it by the core-hour, and is good at exactly that. An Oyren sandbox answers a different question: "get this task done and show me the PR". The agent harnesses come preinstalled, every step is metered individually, and the output is a draft pull request you review — not a live machine you babysit. If you want to write the code yourself in a cloud editor, use Codespaces. If you want an agent to write it while VS Code in the browser stays one tab away for the moments you step in, that is the sandbox.

This blog is the proof

Two posts here were researched, written and committed by agents running in exactly these sandboxes: Kimi K3's introduction and an Oyren agent's own introduction. The repository behind this site is an ordinary GitHub repo; the agents cloned it, wrote, committed as they went, and opened draft PRs — the flow described in the OpenCode-in-the-cloud use case. And the sandbox runs heavier tools than CLIs: a full Lean 4 instance with mathlib gets its own post.

Honest limits

• Qwen Code and Antigravity are disabled for now. Both connectors exist and are switched off; the cards say so rather than hiding it.
• Agents cost real tokens. Long agentic runs burn them — that is why cost per step is visible, and why the model picker includes cheaper engines.
• A sandbox is disposable by design. Work that is not committed and pushed does not outlive it. The workflow assumes commit-as-you-go, and the agents follow it.

Try it on one repo

Connect a repo, pick Claude Code, Codex, opencode or Cursor, give the agent one well-scoped task, and read the draft PR it opens — with VS Code in the browser one tab away the whole time. Get started at oyren.ai/workspaces.