Skip to main content
An AI agent that can run terminal commands can use the SiteGPT CLI to set up and manage chatbots for you. This page shows how to give the agent its own profile and the CLI instructions. If your AI app supports MCP but cannot run terminal commands, use MCP instead. If you do not have a SiteGPT account yet, the agent can use agent-first onboarding and needs no login.

Before you start

  • A SiteGPT account on the Growth plan or above. See Plans and limits.
  • An AI agent that can run terminal commands on your computer.
  • Node.js 18 or later, with npm, on the same computer.
  • A decision about what the agent may do. Pick a scope set in Scope sets for common jobs.

Set up the agent

1

Install the CLI

2

Log in with a profile for the agent

Use a separate profile, so you can see when the agent works and revoke its access alone.
To limit the agent, add --scope and --chatbot options. See Login options.
3

Approve access in the browser

Open the URL that the CLI prints. On the approval page, check the access that the CLI asks for. You can narrow it there. Then select Approve.
4

Give the agent the CLI skill file

Tell the agent to read this file before it uses SiteGPT:
Some AI products call this a skill, project instruction, custom instruction, or memory. To install the skill as a package instead, see Install the skill as a package.
5

Paste the agent instruction

Copy the agent instruction into the agent’s instructions or chat. If you used a different profile name, change sitegpt-agent in the text.

Check that it works

Ask the agent: “Run sitegpt --profile sitegpt-agent whoami and list my SiteGPT chatbots as JSON.” The agent shows your account email and profile name. Then it lists the chatbots that the token can access.

Agent instruction

Example requests

Try SiteGPT before signup. Playbook: New customer onboarding.
Create a chatbot in your account. Playbook: Existing account setup.
Add and check content.
Review recent support activity.

Keep the agent safe

  • Give each agent its own profile, such as setup-agent, knowledge-agent, or support-agent.
  • Limit the token to the chatbots the agent needs with --chatbot.
  • Use --full-access only for a trusted local agent that must manage the whole account.
  • Ask the agent to run sitegpt profiles list and sitegpt whoami before it makes changes.
  • Revoke or rotate the agent’s token after demos and shared sessions. See Token commands.

If the agent has problems

For more fixes, see Troubleshooting.

Reference

Install the skill as a package

The sitegpt/agent-skills repository has two skills:
  • sitegpt-cli: the CLI workflow and command map. It is a packaged copy of the public skill file. The public file can be a little newer.
  • sitegpt: a short orientation skill. It tells the agent which surface to use (CLI, MCP, REST API, or no-account onboarding) and sends CLI work to sitegpt-cli.
The repository is packaged as an Agent Skills repository, a Claude Code plugin marketplace, and a Cursor plugin. It also has config files for Cursor, Windsurf, and Claude Code. The skills only teach the agent. The agent still needs the CLI from Install the CLI.
In Cursor, run /add-plugin and select SiteGPT. The Cursor plugin also adds the MCP server. The CLI also prints a short version of the agent workflow and command map with sitegpt agent-guide. Add --json for JSON output.

Agent discovery reference

These public URLs help agents find SiteGPT without earlier instructions.

Sign in with AgentID

An agent with its own AgentID identity can create a full SiteGPT account without a person. AgentID identities use an AgentMail inbox.