Skip to content
AOZUMIDocs

Cloud agents and plans

A cloud agent is a sandbox of its own on AOZUMI that runs Codex or Claude Code for your threads. You create and manage agents in the Cloud console or in Hachi, and talk to them in Hachi.

Get access

Cloud agents are in a private beta, by invite.

  1. Open AgentsIn the Cloud console, select Agents in the sidebar. Without an invite it carries a Beta tag.
  2. Join the waitlistA workspace owner or admin can add a note, “What would you hand it first?”, and select Join the waitlist.
  3. Wait for your turnThe AOZUMI team lets workspaces in a few at a time. The page unlocks on its own.

A workspace has one open request at a time. If a request is declined, an owner or admin can ask again 7 days later. Until then, nobody in the workspace can create, start or wake cloud agents.

Create an agent

Hachi creates one for each project on its first cloud thread, so you rarely need to. To make one yourself:

  1. Select New agentOn the console's Agents page, or in Hachi's Settings › Cloud agents.
  2. Name it and pick a harnessCodex signs in with ChatGPT or uses the workspace's OpenAI key. Claude Code runs on the workspace's Anthropic key.
  3. Select Create agentA Codex agent that has no OpenAI key to use offers Sign in with ChatGPT: enter the one-time code it shows on OpenAI's page.

In Hachi, the composer's agent picker sends a thread to any agent, or to New agent….

Start, stop and delete

Each agent row shows its harness, its sign-in state, its size, when it was last active and whether it is running. Its menu has:

  • Start and Stop. A stopped agent keeps its files and history.
  • Size, on plans with more than one size, while the agent is stopped.
  • Sign in with ChatGPT and Sign out of ChatGPT, for Codex agents.
  • Delete agent, which removes the sandbox and its history. Hachi lists the threads that ran in it before you confirm.

You rarely need Start: sending a message to a sleeping agent wakes it, and a cold start takes about 20 to 30 seconds. AOZUMI stops an agent after it has been idle for your plan's idle time.

Model keys

A workspace has one OpenAI key and one Anthropic key, shared by every agent in it. Add them under Model keys on the console's Agents page, or in Hachi. AOZUMI encrypts each key and never shows it again; you can replace or remove it.

If an OPENAI_API_KEY or ANTHROPIC_API_KEY is set on your Mac, Hachi offers Add as workspace key. It sends the key only after you agree.

Plans and agent hours

Your workspace's plan sets how many agents it can have, how many run at once, and how many agent hours it gets a month. The Agents page and Hachi's Settings › Cloud agents show your plan, what you have used and when it resets.

PlanAgentsRunningHours a monthSizesIdle stop
Free3125Small10 min
Plus103120Standard20 min
Pro2010600Standard, Large45 min
SizeMachineOne hour counts as
Small0.5 vCPU, 2 GB memory0.5 agent hours
Standard1 vCPU, 4 GB memory1 agent hour
Large2 vCPU, 8 GB memory2 agent hours
  • Agent hours count only while an agent runs, weighted by its size as above.
  • They reset monthly, on the first of the month (UTC).
  • At the limit, agents stop. Nothing runs past it, and agents start again after the reset. Files and history are kept.
  • Cloud standby uses hours, because the cloud agent stays running while your local session works.

Workspace alerts can tell Slack or a webhook when Agent hours 80% used and Agent hours used up.

Archiving and deleting threads

  • Archive ends the thread's session. Sending a message to it later wakes the same conversation.
  • Delete removes the thread's own folder from the agent.

Neither stops nor deletes the agent, because other threads may use it. AOZUMI stops idle agents on its own.

Questions

What happens when my workspace runs out of agent hours?
Running agents stop, and agents cannot start again until the hours reset at the start of the next month (UTC). Their files and history are kept.
Does a stopped agent use agent hours?
No. Agent hours count only while an agent runs. A stopped or sleeping agent keeps its files and history and uses none.
Who can create agents?
Workspace owners, admins and builders can create agents, add model keys and sign an agent in to ChatGPT. Viewers can see agents but not change them.

Checked against the product on .