agent-cli-tools

Agent CLI Tools

Start Here

Use the installed command itself as the source of truth before relying on remembered flags:

where.exe codex
where.exe claude
where.exe kimi
where.exe grok
codex --help
claude --help
kimi --help
grok --help

On Windows, refresh PATH for a new-terminal view when needed:

$env:Path = [Environment]::GetEnvironmentVariable('Path','User') + ';' + [Environment]::GetEnvironmentVariable('Path','Machine')

Read references/installation.md when a command is missing, the user asks how to install/update, authentication is stale, PATH looks wrong, or exact local paths/versions matter.

Choose The Agent

• Use codex for OpenAI Codex local repository work, Codex review, AGENTS.md, skills/plugins/MCP, codex exec automation, and Codex cloud handoff.
• Use claude for Anthropic Claude Code tasks or when the user explicitly wants Claude's coding agent.
• Use kimi for Kimi Code agent tasks, Kimi API Platform or Kimi managed login, kimi web, kimi acp, or Kimi provider experiments.
• Use grok or agent for xAI Grok Build, Grok model testing, or tasks the user wants routed through Grok.

Prefer the agent the user names. If the request is comparative, run the same small prompt against each logged-in agent and report exact command/output differences.

Common Workflows

Validate the local setup

Run versions first, then one tiny prompt:

claude --version
kimi --version
grok --version
claude -p "Reply with exactly: ok"
kimi -p "Reply with exactly: ok"
grok -p "Reply with exactly: ok" --no-alt-screen

Kimi may print extra session commentary after the model answer. Grok may print non-fatal telemetry warnings after a successful answer. Treat the model answer and exit code as the primary signal.

For review or editing tasks, a tiny prompt is only an authentication smoke test. Do not count an agent run as a completed review unless it returns actionable findings, a concrete edit, or an exact no-change verdict for the artifact under review.

Use Codex CLI

If codex is installed, start it in the project directory:

cd C:\path\to\repo
codex

Useful interactive slash commands from the official CLI entrypoint include:

• /init to scaffold an AGENTS.md.
• /status to inspect session configuration.
• /permissions to review or change allowed actions.
• /model to choose model and reasoning effort.
• /review to review changes.

Use non-interactive or workflow-oriented commands only after checking codex --help in the current environment:

codex exec "Explain the repository structure and stop before editing."
codex resume
codex cloud
codex mcp
codex completion

Use codex --search when the task depends on current external documentation, and codex --image when the first prompt needs visual context.

Use Claude Code

claude
claude -p "Summarize this repository without editing files."
claude auth status
claude auth login

Use claude -p for one-shot automation. Use the interactive TUI for multi-step code work where command approvals, diffs, and follow-ups matter.

Use Kimi Code

kimi
kimi -p "Summarize this repository without editing files."
kimi doctor
kimi login
kimi provider list
kimi web
kimi acp

Use kimi doctor before debugging providers or config. Use kimi login for the device-code flow. Avoid kimi -y or --auto unless the user explicitly asks for auto-approval.

Use Grok Build

grok
grok -p "Summarize this repository without editing files." --no-alt-screen
grok login --device-code
agent --version

agent.exe is an alias installed with Grok Build on this machine. This Grok version does not expose grok auth status; validate auth with grok -p "Reply with exactly: ok" --no-alt-screen.

When Grok can answer tiny prompts but fails a real repository review, diagnose the failure before retrying the same shape:

grok --version
grok update --check --json
grok --help
grok -p "Read README.md and print the first heading only. Do not narrate." --no-alt-screen --max-turns 4

If the run logs tool_output_error, read_file failures, only planning narration, or max turns reached without findings, do not count it as a usable review. Retry once with a prompt file that embeds the relevant artifact text, because -p/--single expects the prompt as an argument and piping text into it does not supply the prompt. If the prompt-file run still fails to produce findings or an explicit no-change verdict, switch to another approved agent such as Claude or Kimi and report the Grok failure precisely.

Authentication

Prefer browser/device-code flows over handling raw API keys. The user's preference is to complete browser login in the sidebar or in-app browser rather than the main/system browser when possible.

Do not inspect browser cookies, local storage, password managers, or token files. Do not print API keys or full auth tokens. If an env var is needed, report only whether it is present:

$names = 'ANTHROPIC_API_KEY','MOONSHOT_API_KEY','KIMI_API_KEY','XAI_API_KEY','GROK_DEPLOYMENT_KEY'
foreach ($name in $names) {
  $v = [Environment]::GetEnvironmentVariable($name,'Process')
  "$name=" + ($(if ($v) {'set'} else {'missing'}))
}

Safety

Run agents from the intended working directory and check git status before allowing edits in a repo. Avoid auto-approval flags such as --dangerously-skip-permissions, --allow-dangerously-skip-permissions, kimi -y, kimi --auto, grok --always-approve, or broad bypass permission modes unless the user explicitly requests them for a trusted workspace.

For public repositories, keep private strategy, credentials, account details, and API keys out of files and logs.