Use one remote MCP connection named ai-accelerator, at the aggregate /admin/mcp resource advertised by your installation. The server exposes Insight, Agents and Platform as tool catalogues through that connection. You do not need a local MCP server or the Admin Kit just to connect.
Connect directly
- On the product site, choose Connect your AI assistant and copy its connection prompt into Codex, Claude Code, Cursor or another client with native HTTP MCP OAuth support. The prompt obtains the actual resource URL and issuer from that installation; do not construct them from a guessed domain.
- Let the client open the browser and sign in with your existing admin account.
- Select one or more organizations. Selecting a parent alone does not include its children. Choose Include descendants explicitly to include authorized current and future descendants.
- Select the scopes and maximum role, review and confirm. The client receives short-lived access and rotating refresh tokens, not a new API key. It manages those credentials in its own OAuth store.
For Codex, after configuring the URL, the authorization command is:
codex mcp login ai-accelerator --scopes admin:read,admin:write,mcp:insight,mcp:agents,mcp:platform
Requesting these scopes offers the choices in consent; it does not grant them automatically. Claude Code and Cursor use their native MCP authentication controls. If a client lacks OAuth support, explain that limitation; do not silently create an API key.
Choose access
| Choice | Effect |
|---|---|
mcp:insight | Opens the inspection catalogue |
mcp:agents | Opens the agent/workflow catalogue |
mcp:platform | Opens the platform-resource catalogue |
admin:read | Allows read operations within the granted access |
admin:write | Allows changes within the role and granted access |
| Viewer | Read-only role ceiling |
| Admin | Creation and updates where allowed |
| Owner | Includes deletion where allowed; resource protections still apply |
Consent defaults to read-only. Select admin:write first to make the higher role choices available. Your own organization access also limits those choices. Selecting every scope or the System organization does not grant global superadmin authority. Global operator operations can remain unavailable. An empty search or an unavailable-resource 404 does not prove that a resource has been deleted.
The effective permission is the intersection of current user access, selected organizations and descendant inclusion, scopes, role ceiling, resource ownership and applicable product entitlements. The server advertises only the tool/actions available to that connection; a static reference is not proof of permission.
Verify and continue
The local browser callback and its “Authentication complete” message belong to the client. The client chooses the callback address and port; there is no fixed customer callback port to configure in the platform.
After login, continue the original task: refresh MCP discovery through the client's supported controls, list available tools, and run one small read-only query in an approved organization. For a controlled acceptance run, use the installation's existing Test tenant. Never create, update or delete data solely to prove access. Report:
- Whether OAuth, discovery and the read call each succeeded.
- Which organization was actually checked and which catalogues/actions are available.
- Whether changes are authorized, without claiming that unexecuted writes were tested.
- Two or three useful next requests based on the actual tool inventory.
An already-open chat can retain its previous tool catalogue. Reload its MCP connection or restart the client as supported, then continue. If tools still cannot load, report the remaining step and leave verification pending. The callback alone is not an end-to-end result.
Reauthorize, revoke and migrate
To expand or change access, run the client's authorization flow again and explicitly approve the new organizations, scopes and role. Restarting does not expand a grant. Revoke the connection from Connected applications in the admin; uninstalling the Kit does not revoke OAuth.
An earlier API-key connection is independent of OAuth. Its permissions do not change when OAuth succeeds. Review user-level and project-level client configuration for duplicate connections. Remove or disable only the old entries for this installation after confirming the replacement works; preserve development servers, other installations and other services. Back up configuration first and never print credentials. Existing internal MCP setup is outside this external onboarding procedure.
API keys remain an explicit option for unattended automation or clients that cannot use OAuth. They require their own scopes, role and organization access; a superadmin key does not bypass scope checks. Use secret storage or the documented hidden input, never credentials in prompts or command arguments. OAuth failures are not a reason to switch credentials to bypass a denial.
When to use the Admin Kit
The Kit is an optional project companion: it configures supported clients, adds managed instructions and copies reference docs. Its OAuth setup uses the same connection and consent flow as direct connection. It does not run an MCP server, create an API key during browser authorization, grant extra privileges or make every operation globally available.
The separate installer authorization downloads a signed product release and checks entitlement for one selected organization. It is not the MCP client's grant. Installing or updating the program and authorizing the MCP client are separate steps. The Kit's status output for OAuth describes saved configuration; verify actual tool execution in the AI client.
Project docs already customized by the user are preserved during update-config. Compare them against the installed package's current docs and remove outdated API-key/three-connection instructions manually when appropriate. Treat automatic configuration backups as sensitive if they contain old credentials.