Ongoing discussion: How do we estimate token cost for mcp requests and responses

[brettimus]

Long talk in the office with @flenter about estimating token cost. Here are the notes

Certain mcp methods have a relatively straightforward cost ("tokens") associated with them. For instance:

• tools/call request: the LLM had to produce tokens to call the tool. this token count is of course dependent on the LLM
• tools/call response: the LLM will read the tokens returned in the content block(s)
• resources/read response: the LLM will add the document to context
• prompts/get response: the prompt result will be passed to the LLM

Other mcp methods clearly have no cost associated with them:

• ping
• initialize
• notifications/initialized
• prompts/list req/res: the list of prompts is requested by the mcp client, and stored as application-level state. not injected into the llm's context
• prompts/get request: this is executed by the client, but not the llm

Other mcp methods are ambiguous:

• resources/list req/res: this could be initiated at the application level, or (in the instance of Claude Code) the LLM might have a tool to read MCP resources
• resources/read request: same as above