Skip to main content
The fastest way to give an agent HITL is to let it set itself up. Copy the setup prompt, paste it into the agent, and answer its questions.

​
1. Copy the prompt

In my.hitl.sh, Overview → Copy setup for your agent. The copied prompt includes your API key; on screen it is masked. See what gets copied shows the text first.
The setup prompt expanded under Copy setup for your agent: Connect to HITL.sh for me so you can ask real people for decisions, the masked API key, and numbered steps to connect tools and work with workflows
Without an account, copy it from hitl.sh instead: that version has no key and tells the agent how to sign you up.

​
2. Paste it into your agent

The prompt tells the agent to:
  1. Get an API key. Use yours, or sign you up with your email: npx -y @hitl-sh/cli signup --email <your email>, then ask you for the 6-digit code and run npx -y @hitl-sh/cli signup confirm <code>. It keeps the key in HITL_API_KEY.
  2. Connect a tool, one of:
    • MCP, local: npx -y @hitl-sh/cli init claude-code (or claude-desktop, cursor), or claude mcp add hitl -e HITL_API_KEY=$HITL_API_KEY -- npx -y @hitl-sh/mcp
    • MCP, remote: claude mcp add --transport http hitl https://mcp.hitl.sh/mcp, then /mcp to sign in
    • CLI: npx -y @hitl-sh/cli --help
    • REST: https://api.hitl.sh/v1/api with Authorization: Bearer $HITL_API_KEY
  3. Work with workflows: list_workflows, get_workflow_input, run_workflow, then get_execution or watch_execution until a person answers. To build one: create_workflow, invite_to_workflow (or create_workflow_join_code), activate_workflow.
  4. Read https://docs.hitl.sh/llms.txt for the full reference.
It ends by listing your workflows to confirm the connection and, if you have none, offering to build a simple approval workflow with you as the reviewer.

​
3. Use it

Ask in plain words:
The agent runs the workflow, you get the question on your phone, and the agent continues with your answer. What that looks like in an MCP client: MCP.
Name the workflow in your request, and tell the agent what to do if a person says no. The answer comes back as data (result, decisions), so the agent can branch on it.