What a cloud agent is

A coding agent such as Claude Code or Codex reads a repository, plans a change, edits files, runs commands and tests, and opens a pull request. Run on a laptop, it is a very capable pair programmer that needs a person in the chair. Run as a cloud agent, the same tool works in a remote sandbox: the repository is cloned there, the model calls go out from there, and the result comes back as a pull request instead of a terminal session.

Nothing about the agent itself changes. What changes is who starts it, how many run at once, and what happens when it wants to do something you would rather see first.

Why teams move agents to the cloud

  • Parallel work. One person can have ten runs in flight: a refactor, three bug fixes, a test suite, a review. A laptop holds one.
  • Work that starts itself. An issue handed to an agent, or a step in a CI pipeline, can start a run. Nobody has to sit at a terminal for it.
  • A clean environment every time. Each run gets a fresh sandbox with the repository and tools it needs, so the result does not depend on whose machine it ran on.
  • A record. A run in the cloud can be followed, inspected and kept. A terminal session cannot.

What goes wrong without a control plane

Most cloud agent products give you one agent from one vendor and a place to watch it. That is fine for one person. For a team it leaves four questions open.

  • What will this cost? Model usage is metered by the token, and a run that loops can spend a night’s worth of budget before anyone looks.
  • What if it touches production? Agents run commands. Some of those commands migrate databases or push to main.
  • What did it actually do? A pull request shows the result, not the decisions along the way.
  • Whose keys is it using? Long-lived provider keys copied onto machines are the usual answer, and the usual audit finding.

How AOSplane runs cloud agents

AOSplane is a control plane for cloud coding agents. It is not another agent. It is the loop around your coding agents.

  • One sandbox per run. Each run works in its own isolated sandbox with the repository attached. When the work is done, the run publishes a pull request.
  • Your provider accounts. Model calls go to your own Anthropic, OpenAI, Amazon Bedrock, Azure OpenAI or other provider account with no markup. A credential is handed to a run when it starts, one per run, rather than stored in the sandbox.
  • Budgets checked before a run starts. Organization, project and agent limits are checked together and the tightest one wins. A run whose reservation would exceed the headroom does not start.
  • Approval gates. When an agent asks permission for an action, the request waits in the console with its context. You approve or deny, and the decision goes on the record.
  • A tamper-evident audit trail. Task created, budget reserved, permission decided, pull request handed off: every run leaves a record that is verified server-side.
  • Intake from where the team already works. Start runs from the console, a Linear or Jira issue, the AOSplane agent inside Linear, or the API. Slack is rolling out too: mention @AOSplane or use /aos in a channel. The run stays linked to the issue it came from, and the change comes back as a pull request on GitHub.

Which agents run as cloud agents on AOSplane

Claude Code and Codex run on AOSplane today. Kiro, Cursor and Tower, AOSplane’s own agent, run in preview and might not be enabled for every organization yet. Pick one agent per run, or a different one per role: Claude Code to build, Codex to test, Claude Code to review. Same sandbox, same budget, same gate.

Guides for Claude Code, Codex and Kiro.

What it costs

Platform plans are flat. Model usage is billed by your own provider, with no markup from AOSplane.

  • Free, $0. 2 projects, 2 concurrent runs per project, small machines, 7-day audit retention, $5 monthly usage credit.
  • Starter, $39 a month. 5 projects, 8 concurrent runs per project, small and medium machines, 30-day audit retention, $25 monthly usage credit.
  • Team, $149 a month. 25 projects, 32 concurrent runs per project, all machine sizes, 90-day audit retention, $100 monthly usage credit.
  • Enterprise. Custom limits, invoice billing, 365-day audit retention.

The Free plan is self-serve; paid plans are onboarded with the team. A run that would cross the Free plan’s ceiling does not start, so the free tier cannot spend money you did not give it. Full pricing.

Cloud agents, compared

Two products people weigh against AOSplane: Cursor Cloud Agents and Warp cloud agents. Both run one vendor’s agent on that vendor’s terms. The comparisons are dated and sourced.

Start with one agent

  1. Connect the provider account you already use.
  2. Define an agent: a repository, a harness, a model and instructions.
  3. Hand it a task and answer the approvals as they come.

Grow into a fleet from there. Open the console or read how the loop works.