Connect your own tools with MCP

MCP lets your assistant use your own systems while it chats. It can check stock, look up an order, book an appointment or read a customer's account, and it answers with real data instead of "please contact us".

Looking for an example for your business? The MCP guide has a real use case, a sample conversation and a free starter program for each business type, plus the services you can connect today.

What MCP is

MCP (Model Context Protocol) is an open standard for connecting AI assistants to other software. An MCP server is a small service that offers a list of tools, such as check_stock or find_order. Services such as Stripe, Shopify, Zapier and Airtable run one you can connect, and a developer can build one for any system you run.

You connect the server once. MagizAI reads its tools, you choose which ones the assistant may use and how carefully, and from then on the assistant calls them when a visitor's question needs them.

What you need before you start

• A plan that includes MCP servers. The MCP page shows how many your plan allows. If it says your plan does not include them, upgrade under Billing.
• The server's address. It must be public and start with https://. Servers that only run on a local computer (called "stdio" servers) cannot be connected.
• Its sign-in details, if it needs any: a bearer token, a key sent in a custom header, or a username and password (used by WordPress application passwords and Atlassian). They are stored encrypted and never shown again. Services that only offer an OAuth login cannot be connected yet.

Connect a server

1. Go to Skills and open MCP servers.
2. Under Connect a server, give it a name you will recognise (for example "Inventory system"), paste its address, and pick how it signs in.
3. Press Connect and read tools. MagizAI connects, checks the protocol version and lists every tool the server offers.

If it cannot connect, the page says why. See Troubleshooting below.

Choose the tools and how each is treated

Open the server with Manage. Every tool is listed with the description the server gives it. For each one:

• Use: tick the tools the assistant may call. Anything unticked is never used.
• Treat as Safe: the tool runs straight away. Use this for look-ups that change nothing, such as checking stock or opening hours.
• Treat as Sensitive: before the tool runs, the assistant shows a six-digit code and the visitor has to type it back to confirm. Use this for anything that changes something, such as booking, cancelling or updating an address.
• Filled from the signed-in visitor: if a tool needs the customer's account ID or email, link that argument to the visitor's verified sign-in. The value is then taken from who they really are, never from what they type, so nobody can look up another person's account by typing their email.

Press Save tool choices. Safe tools are ready immediately. Sensitive tools start switched off: turn each one on under Skills when you are happy with it.

Finally, ask your assistant a question that needs the tool, from the chat on your own site, and check the answer. Recent tool calls on the server page shows every call, how long it took and whether it worked.

Signed-in customers

"Filled from the signed-in visitor" needs your website to tell MagizAI who is signed in. Open your chatbot, find Signed visitor identity and generate a secret. Your website's server then signs a short token for each signed-in customer and adds it to the chat widget. The MCP guide has the few lines of code this takes. WHMCS sites get it from the MagizAI WHMCS module.

If a visitor is not signed in, a tool tied to the sign-in is not run, and the assistant asks them to sign in instead.

Limits

• 10 seconds per tool call. The visitor is waiting, so a slower tool is stopped and the assistant answers without it.
• The assistant reads the first 1,500 characters of each result. Short, factual results work best.
• Up to 50 tools per server.
• Simple arguments only: text, numbers, true/false values, fixed choices and simple lists.

How it stays safe

• Nothing is on by default. The assistant can only use tools you tick.
• Destructive tools cannot be marked safe. If the server says a tool is destructive, it always needs the visitor's code.
• Changes need the customer's go-ahead. A sensitive tool waits until the visitor types back the code the assistant shows them. Codes last ten minutes and work once, for that one action.
• Accounts come from the sign-in, not the chat. See "Filled from the signed-in visitor" above.
• Changes on the server are checked. If the server changes or removes a tool, MagizAI switches that skill off and marks it for review. It never quietly starts using a tool that is now different.
• Results are treated as data. Whatever a tool returns is read as information to answer from, never as instructions the assistant must follow.
• Only public, secure addresses. Connections go over https to public addresses only, and each answer is limited to 1 MB.

Troubleshooting

• "The server must be reachable on the public internet over https." The address is private, local, or not https. Put the server behind a public https address.
• "Your plan does not include MCP servers." Upgrade under Billing. Existing settings are kept if a plan changes, and switch back on when the plan allows them again.
• A tool shows "Cannot be used". Its settings use a form MagizAI does not support yet. Ask whoever built the server to keep that tool's arguments to plain text, numbers, true/false values and simple lists.
• "Changed on the server, review and save to use it again". Open the server, check the tool, and save.
• The connection fails with 401. The token in MagizAI does not match the server's. For our starter programs, it must equal the token in config.php.
• The assistant says the visitor needs to sign in. The tool is tied to the sign-in and the visitor is not signed in, or your website's identity token did not verify. Check Signed visitor identity on the chatbot.
• The assistant never uses a tool. Check the tool is ticked, its skill is on under Skills, and its description on the server says clearly when it should be used. The assistant picks tools by reading those descriptions.

For developers building a server

• Start from one of our starter programs on the MCP guide: plain PHP, no dependencies, runs on shared hosting. Or use an official MCP SDK (TypeScript, Python, Go, PHP and others) with the Streamable HTTP transport behind https.
• Write a one-line description for every tool saying exactly when to use it. It is what the assistant reads to decide.
• Mark read-only tools with readOnlyHint and anything that changes data with destructiveHint. MagizAI shows both to the account owner.
• Keep arguments simple: strings, numbers, booleans, enums and arrays of those. oneOf, $ref, pattern and arrays of objects are not supported yet.
• Return short, factual results. Error messages from a tool are never shown to the customer word for word.