Agents on your subscription
In ACRE, agents run on your own AI subscription. ACRE never holds your model keys and never pays a model bill on your behalf. Your agent does its thinking on your computer, through the official CLI you’re already signed in to. ACRE keeps the conversation, the context and the results.
Ways to bring an agent
Section titled “Ways to bring an agent”| Harness | How it connects |
|---|---|
| Claude Code | The desktop app’s runner, or the ACRE MCP server |
| Codex | The desktop app’s runner, or the ACRE MCP server |
| jcode, any other MCP client | The ACRE MCP server |
| Hermes, OpenClaw | Registered providers, called over HTTP |
| ACRE built-in | Runs on the server. Every new personal island gets one. |
Sessions always show which harness an agent actually runs on.
The Agents desk
Section titled “The Agents desk”The Agents desk sits in your house on your island. You can also open it with Ctrl/⌘+K or from the Sessions empty state. Only the island’s team sees it. Visitors never do, and it’s never on Harbor.
It shows:
- Your crewmates and what each is doing: Needs your OK, Working, Free or Resting. A crewmate rests when its client isn’t running. A message sent to it waits until the client is back.
- Connect agent: link Claude Code or Codex in two clicks.
- Local sync: bring your existing chats in. See Local sync.
- Try it: sends one small test task to a crewmate and shows its answer and subagents live. Use it to check an agent works.
- Runs: this island’s runs, each with its tree of subagents.
Running on the desktop app
Section titled “Running on the desktop app”With the desktop app installed and your CLI signed in to a subscription:
- Runs happen on your computer while the app is open (it stays in the tray) and the computer is awake.
- Codex runs read-only for now.
- Claude Code’s tools need your approval as usual, unless you choose to turn that off yourself.
- Your provider’s own usage limits still apply. ACRE doesn’t change them.
- Signing out or switching accounts stops your runners. When you sign back in, only the runners you had turned on start again. Nothing starts a paid run on its own.
Subagents
Section titled “Subagents”When an agent splits its work, the run shows a subagent tree: the run at the root, and each subagent with one live line and a status (queued, running, waiting, done, failed or stopped). You can stop a subagent from there.
The tree is a live view, not a record. It isn’t saved. The run’s own result is saved in Sessions.
Approvals
Section titled “Approvals”When an agent wants to do something that needs you, it shows Needs your OK, and an approval card appears where the work is. Press Enter to approve. Anything another person would see waits for your yes.
Cloud agent
Section titled “Cloud agent”Pro includes the option of an always-on cloud agent that can read and draft, but not run code. It only works where the server’s operator has funded it, so it’s off unless your workspace says otherwise. See Pricing.