Coding Agents

Cue CLI connects your Cue agent on Cue OS to supported coding runtimes and can drive other coding agents — Claude Code, Codex, Gemini, and Hermes — from one interface. Each one has its own top-level command:

Each command wraps that agent's own CLI, so the agent still runs its own loop on your machine. Cue handles session targeting, streaming, and the connection to Cue OS.

• Unified control — one CLI to reach the installed engines
• Remote control — an agent on Cue OS can use one of these as its engine on your computer
• Session management — continue, resume, and list sessions per agent
• Real-time streaming — see output as it generates

For a local worker, cue agent create <name> --runtime <type> offers cue, claude_code, codex, gemini, and hermes. OpenClaw connects through its own binding flow instead. These terminal choices are broader than the web cloud picker. See Agent runtime and models.

Installation

Install each agent's CLI separately:

# Claude Code (Anthropic)
npm install -g @anthropic-ai/claude-code

# Codex (OpenAI)
npm install -g @openai/codex

# Gemini (Google)
npm install -g @google/gemini-cli

Hermes uses its own installation and configuration; see the Hermes command reference.

Authentication

Cue OS sign-in connects the worker to your account; it does not sign external engines in. Authenticate on the same computer where the engine runs.

Codex

Use Codex's native login:

codex login

Cue removes OPENAI_API_KEY when launching native Codex, so exporting that key is not the setup route for cue codex or a Codex worker. The default exec transport normally reads your native Codex home; app-server normally uses an agent-specific home. A configured CODEX_HOME can change this.

For an existing running worker, refresh cue agent runtime status <agent> --check --json first. cue agent runtime login <agent> --engine codex signs into the worker's recorded scope; add --device-auth if you need device-code sign-in. For a remote shell, OpenAI recommends device-code login when the browser callback cannot reach that computer. Enable it in your ChatGPT security settings, or have a workspace admin enable it, then use the login command with --device-auth. This remote route has not been walked on a managed Cue Computer; see OpenAI's authentication guide. Finish the sign-in instructions, refresh status, and send a test message. Running login on your laptop does not sign Codex into a cloud computer. The web OpenClaw Sign in with Codex flow saves an account-wide OpenAI connection that Cue can also use; it does not sign this native engine in. See OpenAI Codex model access.

Claude Code

Use Claude Code's own sign-in, or explicitly supply ANTHROPIC_API_KEY for API billing. For an existing running worker, refresh engine status first, then use cue agent runtime login <agent> --engine claude_code for its recorded scope. Cue's own credentials are not Claude Code credentials.

Gemini and Hermes

Gemini accepts GEMINI_API_KEY or its native browser sign-in; run gemini to follow that flow. Hermes uses the model access configured in Hermes. Check each engine's own status before connecting it to a Cue worker.

To let Cue sandboxes use a Claude Code login, create a setup token with claude setup-token and hand it to Cue:

cue claude connect
pbpaste | cue claude connect --stdin
cue claude connect --from-key anthropic:work --json

The token goes to Cue's encrypted credential store; the command never prints it. A computer signed in as the agent also needs permission to use that saved Claude Code credential. The app grant switches do not currently offer that provider, and there is no guided cloud Claude Code setup. Uploading a token does not select the cloud engine; use your own computer for the setup flow above.

For saved Codex accounts and their quota, use cue codex-auth:

cue codex-auth list
cue codex-auth usage
cue codex-auth switch <account>

Running a Prompt

Claude Code

cue claude "What is 2 + 2?"
cue claude --json "Summarize this repo"
cue claude -c "and now add tests"       # continue the last session
cue claude -r <session-id> "continue"   # resume a specific session
cue claude -n "start clean"             # force a new session

cue claude doctor checks that Claude Code can find the on-disk session JSONL for an agent before you rely on resume:

cue claude doctor --json
cue claude doctor --probe

Codex

cue codex "What is 2 + 2?"
cue codex --json "Analyze this codebase"
cue codex -c "keep going"
cue codex -r <session-id> "continue"
cue codex --transport app-server "run this task"

Gemini

cue gemini "What is 2 + 2?"
cue gemini --json "Explain this code"
cue gemini -c "keep going"
cue gemini -r <session-id> "continue"

Shared Options

Claude Code, Codex, and Gemini share the options below. Hermes uses cue hermes chat and its own session options; see its command reference.

cue codex adds --transport <exec\|app-server>.

From Interactive Mode

Inside cue, /claude runs a prompt through Claude Code:

cue
/claude Write a function to parse JSON

Run cue commands --slash for the current slash command list. Codex and Gemini are reached by their top-level commands rather than slash commands.

Using from the Web or Your Phone

The iPhone and Android apps are coming; store review is pending. You can use Cue OS in your phone browser now.

On the web, an agent's runtime settings can point it at your computer and a coding engine. See Agent runtime and models.

Session Management

cue session list <type> reads the local session history for one agent:

cue session list claude 10
cue session list codex 10
cue session list gemini 10
cue session list claude --latest --id-only
cue session history <session-id> 10 --json

Session types are cue, claude, codex, and gemini.

Session Storage

Each agent stores sessions in its own location:

Next Steps