LoopCode CLI: a terminal loop you can inspect.

Interactive prompts and saved sessions

Keep follow-up work in the same conversation when it depends on earlier evidence. Review older assumptions before continuing a task after the repository changes. Conversation IDs come from the session list; the resume option can match an identifying prefix.

loopcode                              # interactive conversation here
loopcode "Explain the checkout flow"  # begin with a request
loopcode -c                           # continue the latest conversation
loopcode sessions                     # list workspace conversations
loopcode -r <conversation-id>          # resume a chosen conversation

Headless output and turn limits

In plain output, the final answer is written to stdout and tool progress to stderr. json provides a final result object; stream-json provides an event stream. The --json shorthand also selects stream-json. Check the process exit status and result fields before treating an unattended run as successful.

--max-turns bounds the run's model-call budget; it does not replace an external process deadline. --append-system-prompt adds instructions, such as a repository's package-manager convention. It is not a substitute for permission controls.

loopcode -p "Explain the module exports. Do not edit files or run commands."
loopcode -p "Explain the module exports. Do not edit files or run commands." \
  --output-format json --max-turns 12
loopcode -p "Explain the module exports. Do not edit files or run commands." \
  --output-format stream-json --max-turns 12

Unattended permissions need deliberate configuration

Interactive work can ask about an action. Headless work cannot obtain a new interactive answer, so permissions not covered by the current policy are denied. A prompt saying 'read only' communicates intent, but your permission policy establishes the tool boundary.

The CLI exposes -y and --dangerously-skip-permissions for broad approval. These are not recommended defaults for a real project or CI job. If evaluating them, use an isolated disposable environment without production credentials. Prefer scoped permissions and explicit success checks for routine automation.

Before wiring a run into CI, test an allowed task, a denied action, a model or network failure, and an exhausted budget. Ensure the consumer distinguishes incomplete work from a completed result.

Account, usage, and updates

login uses a browser approval code. usage shows the plan and consumed limits, shared with desktop and VS Code. update requests the current CLI release. Keep Node.js 22.18 or newer available on PATH and check the current download page if installation needs repair.

loopcode login
loopcode usage
loopcode update
loopcode --help

External tools through MCP

The CLI includes loopcode mcp add, list, get, remove, and login commands for Model Context Protocol servers. Use the command's help to choose a transport and configuration for the server you trust. Check which tools the server exposes before giving the agent access.

MCP servers can have their own network access, credentials, and side effects. A successful connection does not make every exposed action appropriate for every repository.

loopcode mcp list
loopcode mcp --help