MCP client

What it is

The MCP architecture has three participants. The host is the AI application a person uses, such as Claude Code or an IDE. For every MCP server it connects to, the host creates one client, which keeps a dedicated connection to that server. In everyday speech "MCP client" often means the whole application; strictly, it is the per-server connection inside it.

Why it matters

The client sits between the model and the outside world. It decides which tools the model sees, shows the person what is about to happen, and carries the results back. The MCP specification asks clients to prompt for confirmation on sensitive operations, show tool inputs before calling the server, validate results before passing them to the model, apply timeouts and log tool usage for audit.

How it works

1. The person adds a server, by a local command or a remote URL.
2. The host creates a client, which connects, learns the server's version and capabilities, and, for a protected remote server, runs the OAuth sign-in.
3. The client calls tools/list and hands the definitions to the model.
4. When the model emits a tool call, the client sends tools/call and returns the result to the model.

Example

Adding a remote server to Claude Code is one command:

claude mcp add --transport http actionlake https://connect.actionlake.com/mcp

Claude Code creates a client for that URL, signs the person in through the browser, and the server's tools appear in the session.

Common pitfalls

• Auto-approving every call. The OWASP MCP cheat sheet warns against auto-approval, especially when several servers are connected, and asks for full parameters to be shown, not just a tool name.
• Trusting tool annotations. Hints such as "read-only" come from the server; the specification says clients must treat them as untrusted unless the server is trusted.
• Connecting everything. Each server's tools take context space and compete for the model's attention. See tool overload.
• Assuming the client enforces policy. A client can ask; it cannot stop a server holding a broad credential from doing what that credential allows.

How actionlake handles it

We have tested Claude Code and the Claude app with actionlake. Other MCP clients that support remote servers over Streamable HTTP and OAuth dynamic client registration, or a bearer header, should connect the same way; we have not tested them. The connect page has the exact setup for each, marked with what it rests on. The command above is the documented path for Claude Code, and the quickstart walks through it. The client signs in once with OAuth, or presents an actionlake API key as a bearer token; both resolve to the same key on our side. That key's permissions, caps and approval rules are enforced by the server on every call, so the policy holds whichever client the agent uses. The client sees eight tools however many systems are connected.