---
title: "How do I connect my own MCP server to an AI assistant?"
description: "Paste the server's HTTPS URL, add an API key or sign in with OAuth, and its tools show up. How it works in Kazi, and what to check first."
url: "https://kazi.apifunnel.ai/answers/connect-your-own-mcp-server-to-an-ai-assistant"
updated: "2026-11-14"
---

# How do I connect my own MCP server to an AI assistant?

You give the assistant the server's HTTPS address and either an API key or an OAuth sign-in, and its tools become available in your chats. In Kazi you can connect up to 25 of your own remote MCP servers this way, and each one stays private to your account. You can also just ask Kazi to connect one for you in the chat.

## What you need before you start

A remote MCP server, reachable over HTTPS. A local server that only runs on your laptop over stdio won't work from a hosted agent, because the agent isn't on your laptop.

- The server's URL, ending in whatever path the server documents (often /mcp).
- Either an API key the server accepts, or an OAuth login. Kazi discovers OAuth from the server, so if the server supports it you just sign in.
- A decision on access: should the agent only read through this server, or also change things?

## How it works in Kazi

Add the server from your connections, or say "connect my MCP server at https://…" in a chat. Kazi checks the address, stores the credential and indexes the server's tools for your account only. They are never listed for anyone else.

From then on the tools work wherever you use Kazi: on the web, in Kazi Voice on your phone and by text. You don't re-add the server per chat.

Set the connection to Only read or Can update. The platform checks every call against that setting before it runs. See [How do I give an AI agent read-only access to my business apps?](https://kazi.apifunnel.ai/answers/give-an-ai-agent-read-only-access-to-business-apps)

## Where this goes wrong

Most failures are on the server side: an expired key, an OAuth app that doesn't allow the redirect, or a server that times out on long calls. When a tool errors, read the error before you blame the agent. Usually it's the credential.

The other one is tool sprawl. A server with 200 tools costs context on every turn if the agent loads them all up front. We search tools on demand instead, which I wrote about in [How do you stop MCP tool definitions from filling an agent's context window?](https://kazi.apifunnel.ai/answers/stop-mcp-tool-definitions-filling-the-context-window)

## Related questions

- [How do I give an AI agent read-only access to my business apps?](https://kazi.apifunnel.ai/answers/give-an-ai-agent-read-only-access-to-business-apps)
- [How do you stop MCP tool definitions from filling an agent's context window?](https://kazi.apifunnel.ai/answers/stop-mcp-tool-definitions-filling-the-context-window)
- [How can an AI agent call my own API without me writing an integration?](https://kazi.apifunnel.ai/answers/let-an-ai-agent-call-your-own-api)
