Run tasks with the CLI

Run one-shot and interactive tasks with the Test Companion CLI, and use slash commands, keyboard shortcuts, and approvals in the interactive session.

You can give tc a task in two ways. Both ways use the same agent. In both, you describe what you want in plain language, the same way you would ask a colleague.

Global flags

A flag is an option you add after a command to change how it runs, such as --yolo. The following flags work on a one-shot prompt and on most subcommands. To see the flags of a single command, run tc <command> --help.

Run headless in CI

This section applies when a continuous integration (CI) pipeline, such as GitHub Actions or Jenkins, runs tc for you. For an unattended run with machine-readable output, combine --yolo and --json:

tc "create and run a smoke test for the checkout page" --yolo --json

In CI, branch on the exit code to tell a passing run from a failing test or a blocked run.

Drive tc from another AI coding agent

If you use Claude Code or Cursor, you can let that agent run tc for you. Run the following command in your terminal:

tc agent-skill

The command installs the tc skill into Claude Code at ~/.claude/skills/tc/SKILL.md. The skill teaches Claude Code how to run tc headlessly, including the tc run --json output format.

To print the skill file instead of installing it, add --print with the target: --print claude-code, --print cursor, or --print generic. Use the Cursor output to create a Cursor rules file.

Interactive session

In your terminal, run tc with no arguments to start a session. Type a request, such as create test cases for the login flow, and press Enter. The agent responds, and the conversation continues turn by turn. To leave the session, type /exit.

Slash commands

Type / to open the command menu, then keep typing to filter it. The session includes the following built-in commands:

The agent also accepts three prompt commands. Type /newtask to carry the current context into a fresh task. Type /smol to condense the conversation and free up context. Type /newrule to save a rule from the current conversation.

Keyboard shortcuts

Approvals

Before the agent takes a sensitive action, the turn pauses and shows an approval card. Sensitive actions include running a command, editing a file, and calling a Model Context Protocol (MCP) tool. The card offers three choices:

• Approve runs the action one time.
• Approve always remembers your choice for that action.
• Reject skips the action.

To reject with a reason, type a message instead of choosing an option. The agent receives your message as the reason. One-shot runs with --yolo or --json show no card and approve every action. Use the interactive session when you want to approve each action.

Next steps

• CLI command reference: Look up every tc command.
• CLI configuration and troubleshooting: Configure tc, branch on exit codes in CI, and fix common problems.