AI connector (MCP)

Your people already use AI assistants. The AI connector lets each person connect their own AI tool (Claude, GitHub Copilot, ChatGPT or any client that speaks the Model Context Protocol) to the apps they use, through an endpoint hosted by the portal. It is Vision Agent with a different front end, and it keeps Vision’s rules:

1. You, never more. Every call runs as the person who created the token, with their current apps and roles. Change someone’s role and the very next call follows it.
2. Reads, and requests at most. A change is only ever a pending approval request in the app. The AI tool can never approve.
3. Deterministic. Answers come from the apps’ own code; no AI model of ours is involved and nothing is learned through the connector.
4. Off by default. An organization administrator turns it on under AI tools in the portal. Vision’s own switches stop it too.
5. Audited, in the portal and in each app (“Vision Agent via AI tool for <person>”).

Create a token

1. In the portal, open AI tools in the icon rail.
2. Give the token a name, an expiry (1 to 90 days, 30 by default) and a scope: read, or read + propose to let the tool ask for approvals. Optionally limit it to some of your apps.
3. Confirm your password and authenticator code (needed within the last 10 minutes, because the token outlives your session).
4. Copy the token: it is shown once, together with a ready-to-paste configuration. You also get an e-mail about every new token, never containing it.

At most 10 live tokens per person. A password change, an authenticator reset or a role change ends all of a person’s tokens. Service-provider staff can’t create tokens for customer organizations.

Connect your tool

Use the configuration the page shows; the address is your portal’s own /mcp:

{
  "mcpServers": {
    "cloudvision360": {
      "type": "http",
      "url": "https://my.example.com/mcp",
      "headers": { "Authorization": "Bearer cvmcp_…" }
    }
  }
}

• Claude Desktop: Settings → Developer → Edit config, merge the entry into claude_desktop_config.json, and restart.
• Claude Code: claude mcp add --transport http cloudvision360 https://my.example.com/mcp --header "Authorization: Bearer cvmcp_…"
• VS Code (GitHub Copilot, agent mode): add the server to .vscode/mcp.json and keep the token in a password input rather than in the file.
• Any other MCP client: Streamable HTTP transport, the address above and the Authorization header.

Then ask, for example, “What needs my attention today?”, “Which laptops are low on disk?” or, with a read + propose token, “Ask for approval to block public access on that bucket.”

What the tool can do

Tool	What it returns
list_my_apps	Your apps and your role in each
ask	Vision’s answer to a question, including charts as data and a plain-text description
get_signals	The brief: what needs attention, worst first
get_metrics	Key numbers per app and short series
list_dashboards, get_dashboard	Your custom dashboards and their current values, read as you
list_actions, propose_action	With read + propose only: the requests you may make, and creating one (always pending, never executed)

Everything returned is marked as data, never as instructions, and text people typed (dashboard names, notes) is cleaned and kept separate so a prompt hidden in it can’t steer the tool. No secrets are ever returned.

Administer it

Organization administrators see every connected tool (person, name, scope, apps, created, expires, last used), can revoke one or all, and can turn the connector off for everyone at once. Everyone can see and revoke their own tokens. Rate limits apply per token, per person and per organization.