# MCP Authentication and Authorization: Test Your AI Agent By Steven Van · 2026-10-26 A hands-on guide to checking that a private Quickchat Agent accepts the right MCP caller and rejects anonymous and unauthorized ones. Quickchat has published a walkthrough for testing [MCP authentication and authorization]() against a private AI Agent. [Quickchat]() frames the split this way: authentication identifies the caller, and authorization decides which Agent that caller may use. ## What the walkthrough builds The exercise creates a private Agent called Alder Internal Runbook for a fictional support team. It answers an internal question about a failed nightly export from a single synthetic runbook article. A separate Alder Public FAQ Agent demonstrates anonymous access without exposing the runbook. The expected outcomes are: - A permitted teammate can retrieve the export procedure. - An outsider with a signed-in account but no access receives a server rejection. - Anonymous callers are challenged. - Removing the teammate's access stops their existing connection from working. ## Setup steps Step 1 uses manual setup. The Agent gets a Main Prompt that tells it to give concise factual answers, say when it cannot confirm something, and not invent account details, actions, refunds or recovery results. The runbook is added as a custom article in the Knowledge Base, and the test waits for it to show Ready. Custom Actions stay empty, and Human Handoff and Smart Data Gathering stay off, so the test isolates access to the hosted Agent. The post separates three boundaries: authentication (which account is calling), Agent authorization (whether that account may use the Agent) and data and tool permissions (what the Agent can read or change). It also distinguishes a client calling an Agent's hosted MCP server from an outbound MCP connection where the Agent calls another service. ## What you need - An Agent owner, plus a teammate account and an independent account with no role on the Alder Agent. Both test accounts should complete onboarding and keep an Agent of their own. - Three browser profiles, one per identity, and a terminal with Bash or zsh. - Node.js 22.19 or later to reproduce the MCP Inspector commands. MCP is included on **Free**. The full teammate and removal exercise needs a plan with collaborator seats: Professional includes three users including the owner, and Business includes ten. The public FAQ and the test accounts' own Agents can stay on Free. Each generated reply and Simulation run uses AI credits. ## Test kit and results The author built the demonstration locally and tested the HTTP boundary with real OAuth authorization-code exchanges using PKCE on 26 October 2026. They also connected through MCP Inspector 2.12.0 and ran a three-message Simulation dataset. A real MCP question was recorded in the Quickchat AI Inbox, with an answer drawn from the supplied runbook. A downloadable kit contains the knowledge, Main Prompt, MCP descriptions, test messages, a blank matrix and sanitized local receipts. The files are editable references, not a one-click import. Quickchat notes that the results describe this configuration and are not a certification of every MCP security requirement. [![]()]() ![Quickchat]() Quickchat Build and customize AI assistants for customer support, sales, and internal use — no coding required. [View Quickchat →]() [Original source]() [Read MCP Authentication and Authorization: Test Your AI Agent on Creators Toolbox]() --- Canonical source: https://creatorstoolbox.com/blog/quickchat-mcp-authentication-and-authorization-test-your-ai-agent