How to Reduce MCP Token Costs: 9 Fixes for Users and Server Builders

To reduce MCP token costs, load fewer tool definitions, keep tool results small, and stop paying full price for text the model has already seen. In practice that means switching off servers you do not need, using tool search, capping large outputs, and designing servers that return summaries instead of dumps.

Tokens are how AI providers bill and how much a model can read at once. MCP adds tokens in two places: the tool definitions sent to the model, and the results tools send back. This guide gives nine fixes, split between people who use MCP servers and people who build them. Details were checked against vendor docs on 2 November 2026.

Where MCP tokens go

There are two buckets.

Definitions. Each tool's name, description and input schema is usually shown to the model. Anthropic's advanced tool use post describes "58 tools consuming approximately 55K tokens before the conversation even starts." On a pay-per-token API, you pay for that on every request that includes them.

Results. Whatever a tool returns usually goes into the conversation. Anthropic's code execution post gives the example of a meeting transcript that could add "an additional 50,000 tokens" when passed between tools.

If you use a flat-rate plan, tokens may not show on a bill, but they still fill the context window and can hit usage limits sooner. Our explainer on MCP and the context window covers why this also hurts accuracy.

Fixes for people using MCP servers

1. Audit what is loaded

Before changing anything, look. In Claude Code, the MCP docs say the /mcp panel "shows the tool count next to each connected server." The /context command shows what is filling the context window. Other clients have similar views; check your client's docs.

Write down each server, its tool count and whether anyone used it this week.

2. Turn off servers and tools you do not need

This is the cheapest fix. Disconnect servers nobody uses. Inside busy servers, switch off tools that do not fit the role. Where a vendor offers a smaller endpoint, use it. Linear's MCP docs, for example, offer a read-only URL that "only ever exposes read tools."

The tool permission planner can help you choose a minimal tool set per role. Fewer tools also mean fewer ways for an agent to make a mistake.

3. Turn on tool search

Tool search loads definitions only when the model asks for them. Anthropic's tool search docs say it "typically reduces this by over 85 percent, loading only the 3–5 tools Claude needs for a given request."

In Claude Code it is on by default. You can switch to threshold mode with ENABLE_TOOL_SEARCH=auto, which loads tools up front until they reach 10% of the context window. On the Claude API you add a tool search tool and mark other tools with defer_loading: true. Our guide to tool search and code mode has the details.

4. Cap large outputs

Claude Code "displays a warning when any MCP tool output exceeds 10,000 tokens" and sets "the default maximum" at 25,000 tokens, per its docs. You can change that with MAX_MCP_OUTPUT_TOKENS. Raising it lets more data in; lowering it protects your budget. When a text result is too big, Claude Code saves it to a file and passes the path, so the model reads only what it needs.

5. Ask narrower questions

The model calls tools based on your prompt. "List the five most recent unpaid invoices over $1,000" leads to a small query. "Show me all invoices" may pull thousands of rows. Name the date range, the count and the fields you want.

6. Keep sessions focused

Long sessions carry every earlier result forward. When you switch tasks, start a new conversation instead of piling a second job onto the first.

Fixes for teams building agents or servers

7. Keep tool definitions stable for prompt caching

Prompt caching lets an API reuse text it has already processed at a lower price. Anthropic's prompt caching docs say "Cache read tokens are 0.1 times the base input tokens price" for most models, while cache writes cost 1.25 times (5-minute TTL) or 2 times (1-hour TTL).

The order matters: the cache is built from "tools, system, then messages." The docs warn that "Modifying tool definitions (names, descriptions, parameters) invalidates the entire cache." So avoid changing the tool list mid-conversation. The MCP tools specification supports this too: servers "SHOULD return tools in a deterministic order," which "improves LLM prompt cache hit rates."

Tool search also plays well with caching. Anthropic's docs say deferred tools are excluded from the system-prompt prefix, "so prompt caching is preserved."

8. Design servers that return less

If you build MCP servers, this is where you can save your users the most:

• Paginate. Return a page and a cursor, not everything. The spec supports pagination on list operations.
• Filter on the server. Accept parameters like date range, status and limit.
• Summarize. Return counts, totals and top items, with a way to fetch details.
• Use structured output. A tool with an outputSchema returns structuredContent, which is easier to parse than long prose. See structured tool output.
• Write short descriptions. Put key words first. Claude Code cuts upfront tool descriptions at 4,096 characters by default.
• Merge near-duplicate tools. Three tools that differ by one parameter cost three definitions.

Some servers do this already. Google's BigQuery MCP docs say "Query results are limited to a maximum of 3,000 rows."

9. Move heavy data work into code

For agents you build, consider letting the model write code that calls tools and returns only the answer. Anthropic reported its Programmatic Tool Calling feature dropped average usage "from 43,588 to 27,297 tokens, a 37% reduction on complex research tasks." This needs a code sandbox, so weigh the security work against the savings.

What to measure

Do not guess. Track these over a few weeks:

• Input tokens per request, before and after each change.
• Tool definition tokens at session start.
• The largest tool results, by tool name.
• Cache read versus cache write tokens, if your API reports them.
• Task success rate, so savings do not come at the cost of wrong answers.

Our playbook on measuring MCP ROI covers how to set a baseline.

Next step

Run the audit in step 1 this week; it usually shows where the waste is. If you build an MCP server and want it to be cheap for customers to use, our MCP development team designs lean tools from the start.

Frequently asked questions