How to connect an MCP server to your AI agent

Connecting an MCP server to an agent is a four-step loop: choose a server, choose how it runs, configure it, and verify. If you are new to the protocol, read What is MCP? first — this guide assumes you know what a server exposes.

Step 1 — Pick a server

Start from the capability you need, not the server. Want the agent to read a repo? Look for a code or GitHub server. Want web results? A search server. Our MCP directory groups servers by category so you can find candidates quickly, and each entry records its publisher and transport.

Step 2 — Choose a transport

MCP servers run in one of two common ways:

• stdio — the server runs as a local subprocess on the same machine as the agent. Simplest for local development and for servers distributed as a CLI or npm package.
• HTTP / SSE — the server runs remotely and the agent connects over the network. Better for shared, hosted, or team-wide servers.

Local-first tools (filesystem, git) are usually stdio; hosted SaaS connectors are usually HTTP.

Step 3 — Configure the server on the agent

Most agents and platforms read an MCP configuration that lists each server. Conceptually every entry needs:

Keep secrets in environment variables, never hard-coded in the config. If the server needs an API key, provision it first and reference it by variable.

Step 4 — Restart and verify

Restart the agent so it re-reads the configuration and performs the MCP handshake. Then confirm the new tools appear in the agent's available-tools list. If they do not:

• Check the server actually starts (run the stdio command by hand, or curl the URL).
• Check credentials are present in the environment the agent runs in.
• Check transport — a stdio config pointed at a remote server (or vice versa) will silently fail to connect.

A server that starts but exposes no tools usually means missing auth or an unconfigured scope.

Doing this on a platform vs a framework

How much of this you do by hand depends on what is running your agent:

• On a framework or SDK, you own the config file and the process, so you wire MCP yourself. See the best platforms for coding agents for SDK-shaped options.
• On a managed platform, the platform often manages MCP configuration and credentials for you through its UI. How native that experience is varies — compare it on the platforms directory, and see MCP-native vs bolt-on for why it matters.

Next steps

• Give the agent state between runs: How to give an AI agent long-term memory.
• Keep it running: How to run an AI agent 24/7.