Ask Ari and external agent access

AgentReady has two deliberately separate assistant surfaces: Ask Ari inside the app and a planned external Model Context Protocol (MCP) connection for tools such as Claude or Cursor. They do not share the same write authority.

Ask Ari in the app

Ask Ari is the guided interface inside AgentReady. It can:

• explain the current page and the evidence behind a topic project;
• guide you to the appropriate next step;
• navigate to governed destinations inside your workspace; and
• in one bounded pilot, confirm an allowed profile fact after explicit confirmation.

That confirmation is protected by a signed token, same-origin and workspace checks, entitlement and freshness checks, provenance, and undo. Ask Ari has no model-exposed tools for publishing, spend, billing, Shopify schema writes, or declaring work complete.

External agent access

The external workspace MCP surface is read-only by design. When available, it will let a workspace-scoped client read the same governed record you see in the dashboard without giving the client a general write channel.

Rolling out. This section describes the connection contract. Use the endpoint URL and token shown in dashboard settings only when that surface is enabled for your workspace.

Two values come from the dashboard:

• An MCP endpoint hosted at https://app.joinagentready.com/api/…. The exact path is shown in your dashboard settings when available. It speaks streamable HTTP, the transport every current MCP client supports.
• A workspace-scoped token. It authenticates the endpoint and scopes every call to one workspace, yours, and nothing else. You create it, see it, and revoke it from settings. It is shown once at creation and hashed at rest, so a leaked token is inert the moment you revoke it.

You paste both into your client once. From then on your agent talks to AgentReady over that connection.

Set it up in your client

No client is privileged. Use whichever you like. In every block, replace the bracketed placeholders with the endpoint and token from your dashboard settings.

Claude Code

claude mcp add --transport http agentready \
  "https://app.joinagentready.com/api/<mcp-endpoint-from-settings>" \
  --header "Authorization: Bearer <your-workspace-token>"

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "agentready": {
      "type": "http",
      "url": "https://app.joinagentready.com/api/<mcp-endpoint-from-settings>",
      "headers": {
        "Authorization": "Bearer <your-workspace-token>"
      }
    }
  }
}

Cursor

Add this to .cursor/mcp.json in your project (or your global Cursor MCP config):

{
  "mcpServers": {
    "agentready": {
      "url": "https://app.joinagentready.com/api/<mcp-endpoint-from-settings>",
      "headers": {
        "Authorization": "Bearer <your-workspace-token>"
      }
    }
  }
}

Any other MCP client

Point a generic streamable-HTTP MCP client at the same values:

Transport:  Streamable HTTP
URL:        https://app.joinagentready.com/api/<mcp-endpoint-from-settings>
Header:     Authorization: Bearer <your-workspace-token>

What an external agent can do

Read the record. Your agent can pull your business profile, your tracked question set, your latest runs and answers (each carrying the evidence it was scored from), your fix queue with its jobs and skip reasons, and your verification ledger. This is the same data the dashboard shows, read straight from the system of record.

What it cannot do

• It cannot write through the read-only MCP connection. Reading a recommended change is not approval to apply it.
• It cannot publish or apply to surfaces AgentReady manages. Approval stays in the app and is never delegated to the external conversation.
• It cannot outlive a revoke. Your token is visible and revocable from settings, and killing it cuts the connection immediately. Tools that must not exist on this surface are simply absent from it, not hidden behind a permission you might grant by accident.

The distinction is intentional: Ask Ari is a governed in-app guide with one narrow reversible confirmation action; the external agent surface is read-only.

One surface, or two?

If you have used AgentReady on a Shopify store, you may have met the storefront MCP already. They are different surfaces with one job each, and you can keep both:

• The storefront MCP is shopper-facing. It gives an assistant read-only access to your live products, policies, and brand info so it can research and answer questions about your store. It sits alongside Shopify's native shopper MCP and is not how you operate AgentReady.
• The external agent endpoint on this page is owner-facing and read-only. It reads your AgentReady record so your preferred client can help you understand it.
• Ask Ari is the in-app guide for explanation, navigation, and the bounded profile-confirmation pilot described above.

When you connect your agent to operate AgentReady, the endpoint on this page is the one you want.