An MCP client is the application that opens and manages connections to MCP servers on a model's behalf. Claude, ChatGPT, Cursor and VS Code are all MCP clients. The client discovers what tools a server offers, decides which to call, sends the call, and hands the result back to the model.
Why it matters
The client is where most of your practical limits live, and almost nobody reads its documentation before blaming the server.
Anthropic publishes the numbers for Claude: a tool result is capped at roughly 150,000 characters on Claude.ai and Claude Desktop, and at 25,000 tokens in Claude Code, with a 300-second timeout on the hosted surfaces. Claude also supports the 2025-03-26, 2025-06-18 and 2025-11-25 authorization specs, and does not support sampling or resource subscriptions at all. Those are client decisions. The same server behaves differently depending on which one you point at it.
Example
You ask your agent for 100 companies matching an ICP. The server returns all 100. The client truncates the payload at its own ceiling, the model summarises what survived, and you get a confident answer about 60 companies with no error anywhere in the chain.
Nothing failed. The client did what it documents. You just never saw the boundary.
Common mistake
Treating "MCP-compatible" as a single capability. It is a floor, not a guarantee. Two clients that both speak MCP can differ on which protocol primitives they implement, which authorization spec versions they accept, how large a result they will carry, and how long they will wait. Before you promise a workflow to a team, check the client's own docs for the number that will bite you.
How this relates to Signl
Signl is the server, not the client. It speaks Streamable HTTP and authenticates with either OAuth or an API key, so it works with any client that supports remote MCP servers. The Connect page in your dashboard carries setup steps for Claude, ChatGPT, Cursor and VS Code, plus a generic option for anything else that speaks the protocol.
The client-side limits above are why search_signals takes a limit parameter that defaults to 20 rather than returning everything it finds.