MCP is
Dead.

MCP is wildly token inefficient! I've created a custom harness specifically to test MCP token usage vs. putting the MCP "tool calls" in an Agent SDK.

The harnessed agent has one tool call, which is "execute_code". Using code, the agent can discover what is available in the SDK, and it may register more "tools" in the SDK to persist new capabilities. As a result, using "tools" becomes a constant cost on the context.

The first session for building out the new tooling in the SDK is expensive, but once that is done the cost of calling the agent SDK tooling becomes a constant and does not continue to fill the context.

On the other hand, the cost in terms of tokens for MCP is a linear progression since the MCP surface must travel in the prompt on every turn.

I ran several sets of experiments with the harness using both the MCP server and the Agent SDK to validate my results. The links to everything are below.