Cloud devboxes for the agent you already pay for
Not a black box. A devbox.
Run Claude Code, Codex, and OpenCode in isolated cloud devboxes you can watch, steer, SSH into, and verify. Your provider account covers the model. Aether charges only for the machine.
starts from GitHub · Slack · Linear · CLI · API
Demo: a failing test is reproduced, fixed by the agent, tests pass, and a reviewed pull request opens.
A loop you can watch and take over.
The whole devbox is open while the loop runs: every command and diff streams into the thread as it lands, ports become live previews, and your own terminal is right there in the same box. Press esc to interrupt, or let it finish and verify the work yourself. Autonomy that never asks for blind trust.
- Plan mode: the agent proposes, you approve, then it executes
- Queue follow-ups mid-run and the agent folds them in without restarting
- Stop the agent anytime and the devbox stays alive, yours to use
Demo: an agent streams commands and diffs into the task thread while a typed follow-up queues without interrupting it; in a separate pane, the human runs npm test in their own terminal and the suite passes.
AI that never ships its first draft.
Shipped by an agent. Reviewed by an agent. Approved by you.
Demo: Aether opens a pull request, its review agent finds a double-charge bug and requests changes, the author agent pushes a fix, the review approves, and the pull request resolves with passing checks.
Recorded live, unedited: the reviewer traced a double-charge bug through the callers, the author fixed it, and the loop converged in 12 minutes. No human touched it.
Turn on Aether review and every PR the agent opens runs a gauntlet before it reaches you: a review agent tears the diff apart in GitHub, the author fixes and pushes, and the review runs again until the code holds up. What lands in your repo has already survived review, with a video to prove it works.
Seeing is shipping.
Every PR making visual changes arrives with the receipts: a Playwright video of the change actually running, screenshots, green checks, and the full review thread.
I'd bet my on-call shift on this one. Reverted the fix and the new duplicate-delivery test caught the regression. ✓
Your subscription, multiplied.
You already pay for Claude Code, Codex, or OpenCode. On your laptop that buys one agent, one task, your attention, your battery. Point the same subscription at Aether and it becomes a fleet: parallel isolated devboxes, each ready in seconds, running while you sleep, wired into GitHub, Slack, Linear, and Sentry. The model is yours. We're everything around it.
Start tasks where you already work.
brew install the CLI and aether run streams a cloud task into your prompt. Mention Aether in Slack, assign it a Linear issue, or hit the API with a platform key: every surface reaches the same loop.
Everything in the box.
Every PR lands with a screen recording: a real browser clicking through the change the agent just shipped. Watch it work before you read a line of the diff, and hand reviewers the same receipt.
Give a prompt a schedule: hourly, nightly, or every Monday at 9. Triage the backlog while you sleep, chase flaky tests every morning, and wake up to PRs instead of a to-do list.
A review agent reads every diff the moment it lands and leaves real comments in GitHub. Push a fix and it reviews again, until it approves. Answer its comments and it answers back.
Every port the agent opens gets its own URL. Click into the running app and test it like localhost, hot reload included, or hand the link to a teammate. Private by default.
Mention Aether in a thread; progress posts back to it.
Claude Code, Codex, or OpenCode; 40+ models on your keys.
Assign an issue like a teammate; the PR links back.
Production errors become fix tasks with receipts.
Answer from your inbox; the run resumes where it stopped.
brew install, aether run, ssh into any devbox.
Projects, tasks, and automations behind platform keys.
Commit a SKILL.md; every agent that touches the repo follows it.
Don't pay for intelligence twice.
Black-box agents rent you a model at a markup and hide it inside an opaque meter. Aether charges only for compute: the intelligence runs on the subscription you already own.
- VM timebundled
- model inferencebundled, marked up
- networkingbundled
- computemetered, visible
- model inference$0 to us (your own subscription)
- idle time0 credits
Comparison uses Devin Teams list pricing (devin.ai/pricing) as of July 4, 2026.
Kick the tires on real tasks.
- 1 warm slot: instant starts
- Small & Medium devboxes
- Docker on every run
- No overage: upgrade when you hit the cap
A steady lane of agent work.
- 2 warm slots: instant starts
- All sizes, up to 4 perf vCPUs / 16 GB
- Docker on every run
- Credit packs at $0.10 per credit
The everyday fleet for one engineer.
- 3 warm slots: instant starts
- All sizes, up to 4 perf vCPUs / 16 GB
- Docker on every run
- Credit packs at $0.10 per credit
For a fleet that never sits idle.
- 5 warm slots: instant starts
- All sizes, up to 4 perf vCPUs / 16 GB
- Docker on every run
- Credit packs at $0.10 per credit
1 credit = 6 minutes on a Medium devbox. Small = 0.5×, Large = 2×, XL = 4×. Idle and suspended devboxes bill nothing.
Credit packs: 100 for $10, 250 for $25, 500 for $50, 1,000 for $100. Packs expire 12 months after purchase.
Cancel anytime via the billing portal.
Why trust Aether
Security
AES-256-GCM at rest. One Firecracker microVM per task, never reused. Your model runs on your own provider account.
Read the security page →Built by
Pranav Sharan. Previously infrastructure at Apple, Skyflow, and HeyGen. Aether came from wanting cloud agents without surrendering the environment they run in.
About Aether →Receipts
A public demo repo, two films recorded live and unedited, and the complete PR the review loop shipped and re-verified.
Inspect the demo PR →FAQ
Encrypted at rest with AES-256-GCM under a key derived per user, and injected into your isolated VM only for the duration of a session. You can revoke them at any time. The security page describes the full picture.
Something else on your mind? Email us.
Connect a repo. Bring your agent. Go.
01Connect GitHub
Connect a repo.
02Sign in with your Claude Code, Codex, or OpenCode account
Bring your agent.
03Assign a task from Slack, Linear, or the app
Go.
Stop firing agents into the void.
Run your first agent in a devbox you can actually see. Free.
Bring your own subscription. Your agent, your keys.