Since coding agents like Claude Code and Codex came out, I've been pretty obsessed with them. It's hard not to when you're getting a 20x discount on inference. I've also been frustrated by them a lot. I hated walking around with my laptop lid open while connected to a hotspot on my phone, having to tediously accept commands so I don't lose my root directory, and running out of RAM when running multiple agents that were spinning up dev servers in different worktrees at once.

A lot of people would say existing cloud agents are the solution to this, as none of the above problems apply to any cloud agent, but all the cloud agent products I tried had problems worse than local agents. Waiting for the VM to get set up when responding to a task made them feel like they were only for fire and forget workflows, and I missed going back and forth with the agent and using plan mode. On top of that, the whole VM was a black box and I had no way of verifying the agent's work besides checking out its branch and testing its changes myself, which got annoying super quickly.

This is what led me to build Aether, which gives you the best of both worlds. You can start an agent on any repo of any size within 5 seconds, with dependencies installed and dev servers running. On responses the VM is ready in less than a second, so going back and forth with the agent feels like a local session. Plan mode and skills work like they do on your machine.

The VM also isn't a black box. Every workspace is a full devbox with a terminal, file editor, port previews, and Docker. You can step in to steer at any point, or manually test the agent's work through the workspace's exposed ports, just like you would locally.

Running in the cloud also enables a lot of stuff you can't do with a local agent at all. Aether can review PRs with a custom review prompt and leave comments on the PR. Agents can also be set to respond to PR comments, and when both are enabled, it creates a loop between the writer agent and reviewer agent, and the agents keep working until they converge (probably my personal favorite feature). Getting the same result with local agents would be an extremely tedious multi hour chore of copying and pasting between terminal tabs. Here's an example PR of the agents going back and forth in this loop: https://github.com/Aether-Runtime/acme-checkout/pull/4 (it's our public demo repo, the bugs are seeded on purpose, disclosed in the readme)

Other features of Aether include:

- PRs making visual changes come with video demos and screenshots of the changes

- Start agents from Linear issues or Slack channels and communicate with them in the comments/thread

- Set up the Sentry integration with 1 click to start an agent when an alert fires. The agent gets full access to the Sentry MCP, so it pulls the stack trace and context itself

- Set up automations to run scheduled jobs

We run the Firecracker layer ourselves instead of wrapping a sandbox provider, which enables us to be very competitive on pricing. The inference runs on your own subscription, so you still get the huge subsidization and Aether only charges for compute. There's a free tier and paid plans start at $30/mo.

Really appreciate any feedback, positive or negative, and happy to answer any questions!

Show HN: Aether – Run Claude Code, Codex, or OpenCode in devboxes you can watch