Agent runtime and models
Choose where your agent runs, which engine does its work, and which model it uses.
Your agent can run on Cue's servers, on its own cloud computer, or on your computer with Cue CLI. The engine is the software that does its work; the model is the AI that engine calls. Choosing a GPT or Claude model in Cue does not switch the engine to Codex or Claude Code.
New agents created with Cue CLI using Cue or Codex use GPT-6.1 Sol and xhigh thinking as their normal runtime defaults. The quick conversational phase has its own effort setting. Existing agents keep their saved selections.
Where agents run
| Place | Engine choices today | What you need |
|---|---|---|
| Cue's servers | Cue | A Cue OS account. Normal signup and additional agent creation start here, without a computer. |
| Cloud computer | The web setup offers Cue, OpenClaw, and Codex when installed. Codex and Claude Code are included in the computer software; native sign-in still happens on that computer. | A Cue Computer. Cue and Cue-managed OpenClaw use its Cue sign-in; native engines need their own credentials on that computer. |
| Your computer | The web can select Cue, Codex, Claude Code, or an existing OpenClaw agent, according to what the connected worker offers. The picker can also show other engines advertised by the worker, such as Gemini or Hermes; CLI creation offers both. | Cue CLI, a connected agent, and any external engine installed and signed in on that computer. |
Cue on the servers can search the web, use memory, run routines, and use connected services within their granted access. A shell or software-building task can start a cloud computer when needed. You do not need to create a computer just to begin talking.
Creating a local worker with cue agent create is different from creating an account agent in the app: it creates a workspace on the computer where you run the command. Pairing connects that worker to an existing account agent.
On the web
- Open Chat, then Agents, and open your agent's details.
- Choose Computer. If you are already in its workspace, choose Settings → Computer, then the Computer button. Some messages still call this AI Provider.
- Under Where the agent runs, choose Cue's servers, Cloud computer, or Your computer.
- For Cloud computer, create a computer if prompted. The web's Agent engine choices include Cue (default), OpenClaw, and Codex when installed. Selecting OpenClaw starts its setup; see OpenClaw.
- For Your computer, connect an online Cue CLI worker, or choose Show setup steps. Follow the displayed install, sign-in, and one-time pairing commands on your computer. The pairing command connects this agent; the code expires after about ten minutes. The connected worker's engine controls include Codex and Claude Code when available; OpenClaw needs an existing OpenClaw agent name and gateway.
- Send a short message, such as “Reply with the word OK,” and wait for a reply before starting longer work.
Cue CLI has native Codex and Claude Code engines on computers. Cloud Codex has web model and thinking controls after its native login is ready. For native installation and sign-in, use Cue CLI. iPhone test builds can select an engine advertised by an already configured online worker, but do not install or sign it in.
Ask your agent
In your own conversation with your agent, ask “Move to your cloud computer,” “Run on my computer,” or “Move back to Cue's servers.” The agent describes the move and asks you to confirm before applying it. If there is no suitable connected computer for a local move, it gives you a pairing command to run there. Choose the computer engine with the controls above.
These requests keep the same agent and conversation. They do not automatically copy files between your computer and a cloud computer. You can also ask for a local engine change as described below. Use the engine controls for OpenClaw and for cloud engine setup.
Change the engine in chat on your computer
Once your agent is connected to your computer, ask “switch to Codex” or “switch back to Cue,” then confirm when it asks. Your next message goes to that engine on the same computer. The agent keeps its account, conversation, workspace and saved engine settings. Missing Codex model or thinking choices use the current Cue CLI defaults.
Codex must already be installed and signed in on that computer. Cue is included with Cue CLI. Keep Cue CLI running and connected while changing engines. If the computer is offline or engine changes from chat are disabled, the agent explains what needs attention and stays assigned to that computer.
Thinking on Cue and Codex
In Computer, select where the agent runs and its engine. For Cue on either computer, choose Model and Thinking. For Codex, use its Model and Thinking controls after signing native Codex in on that computer.
Quick, Balanced, and Thorough select low, medium, and high effort. A saved web choice applies to the next chat turn, including a short reply. Changing a setting during a turn takes effect on the next one. Higher effort can take longer. Codex's default option clears the web effort override.
The choice is saved for this agent conversation. Cloud Cue also saves the computer's runtime default; Codex saves its native engine setting. A worker's turn logs show the applied effort. A saved picker value or a successful reply alone does not identify the effort used.
On iPhone, Android, and Mac
In test builds, the iPhone Computer screen can return an agent to Cue's servers, bind an already connected online computer, and select an engine that worker advertises under Powered by. OpenClaw must already have a named binding. The screen does not install engines, sign them in, or create a new cloud computer. Model controls differ by engine: server settings and online Codex/Claude Code settings are editable; Cue-computer values are more limited. Availability depends on the app build.
The Android app has no manual computer or engine controls; request a move in the conversation or use the web controls above. Talking from either phone does not run the agent on the phone.
Current Mac development builds offer Settings → Computers → This Mac → Use this Mac to connect an owned account agent using Cue. This prepares a Cue CLI companion when needed. It has been verified in a development build. The public beta download does not yet include it. Desktop Code sessions separately offer Cue, Codex, and Claude Code on the Mac. See Cue Desktop.
Models and sign-in
Cue-managed model use normally draws on Cue credits, billed to the creator even when the computer signs in as the agent. If you connect your own OpenAI Codex sign-in or OpenAI API key, eligible OpenAI requests use that credential instead; an agent account needs its own connection or a grant from its creator. See OpenAI Codex model access for plan limits. External engines use their own credentials and model availability. Signing into Cue OS alone does not sign native Codex or Claude Code in, and local execution can still call a remote model.
For Codex, use its native login on the computer where it runs. Cue removes OPENAI_API_KEY when launching Codex. The web's OpenClaw Sign in with Codex option saves an account-wide OpenAI connection usable by Cue and OpenClaw; it does not sign the native Codex engine in. See Coding agents.
Model lists can differ between Cue, OpenClaw, and each native engine account. Server conversation settings, computer engine defaults, and a model chosen for one CLI run have different scopes. Check the engine and model reported by the connected worker, then send a test message. A worker engine change can affect other conversations handled by that worker.
With Cue CLI
- List your local agents and check one:
cue agent list
cue agent status <agent> --json- See the models you can use, and the details of one:
cue model list --json
cue model show <model-id> --json- Save a model for an agent, or use a model for one run only:
cue agent set-model <model-id> <agent>
cue --agent <agent> --model <model-id> run "Reply with the word OK"- To run a local agent on another coding engine, create it with
--runtime.cue agent create --helplists the engines you can choose. The engine must already be installed and signed in on your computer; see Coding agents.
cue agent create <name> --runtime codex- To see OpenClaw's connection on a Cue Computer, run this in the computer's terminal:
cue openclaw status --jsonLimits today
- Which engines and models are offered depends on the surface, installed software, sign-in, and connected worker.
- Selecting a location, engine, or model does not prove the agent is connected. Send a test message.
- After switching a cloud agent from OpenClaw back to Cue (default), replies can keep coming from OpenClaw. OpenClaw shows how to remove that binding.
- Your computer must stay running and connected for its agent to reply. An offline local worker does not silently move back to Cue's servers.
- Running unattended for weeks at a time has not been proven.