Search MCP Context Cost and Token Use

Where the tokens go

Most MCP clients load every tool definition into the model before the first message. Each name, description and schema counts toward the context. Anthropic's engineering article says that with thousands of connected tools, an agent may have to process hundreds of thousands of tokens before it reads a request.

Each tool result also passes through the model. Full page text or long result lists add tokens on every call, and they stay in the conversation for later turns. The article notes that the model often has to copy intermediate data into its next call, which adds more cost and more room for error.

Context also grows with each turn, so a session that started cheaply can become expensive after many large results. Start a fresh conversation when the task changes, and carry over only the short answer you need.

Result size is the lever

Compare three shapes of output. A ranked list of links is small. Snippets with matching passages are medium. Full page text is the largest. Hyperresearch's search returns passages copied word for word, and its ask returns one short answer with quotes, so the agent reads less text per call.

Some servers let you cap results. Brave's web search takes a count from 1 to 20, and its LLM context tool has controls for token and snippet limits. Check each vendor's parameters before you ask for the largest result set. Anthropic's article also describes an approach where the agent writes code that calls tools and filters data before the model sees it. In one example, the article reports a drop from 150,000 tokens to 2,000.

Keep the cost down

Enable only the tools a task needs. Split a broad research job into focused calls. Ask for a short answer when you need one fact, and fetch a full page only when you must quote more than the answer.

Tool descriptions are part of the bill too. A server with a long list of verbose tools costs more than a short one that offers the same search. When you compare two servers, compare their tool lists as well as their answers, and disable the extras.

Checklist

• Turn off search tools you never use, so their definitions do not load.
• Prefer answer or passage tools over full-page fetches for fact checks.
• Set result counts and token limits where the vendor offers them.
• Split a broad question into several focused calls.
• Check your client's usage display after each call and trim large results.

Questions & answers

Sources for this entry

• Anthropic Engineering: Code execution with MCP
• MCP specification: Tools
• Brave Search MCP server (GitHub)
• Hyperresearch Answers

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

• Servers
• Search MCP Server
• Resource directory

Cite this entry