MCP servers or command-line tools: what each adds to every message

Claude Code is Anthropic's coding tool. You type a request, and an AI model reads files, runs commands and edits code for you. The work is counted in tokens. A token is a piece of text that the model processes, and every request you send is measured in tokens.

Claude Code can reach outside services, such as an issue tracker (a tool that lists a project's bugs and tasks) or a cloud account, in two ways. The first is an MCP server, a program that connects Claude Code to an outside tool or data source. The second is a command-line tool, a program that is run by typing its name in the terminal, which Claude can run by itself.

This page belongs to the guide on how to reduce Claude Code token usage. It compares what each of the two adds to every message. The short answer, from Anthropic's documentation read on 4 October 2026: with default settings an MCP server adds its tool names and its instructions to every request, and a command-line tool adds nothing until Claude runs it [1][2].

What an MCP server is, and what a tool definition is

MCP stands for Model Context Protocol. Anthropic's documentation, read on 4 October 2026, describes it as an open source standard for connecting AI to tools, and says MCP servers give Claude Code access to your tools, databases and other services [1]. The same page says when to connect one: when you find yourself copying data into the chat from another tool, such as an issue tracker or a monitoring dashboard [1].

Each server offers tools. A tool is one action that the model can ask Claude Code to perform, such as "search the issues". Before the model can call a tool, the request must contain the tool's definition. A tool definition is the text that tells the model what the tool does and what inputs it takes.

Definitions matter for cost because of their place in the request. The context window is the text the model can read in one request. The model keeps nothing between two requests, so Claude Code sends the full context again every time [4]. Tool definitions are in the system prompt layer, which is the first part of every request [4]. A definition that loads at the start of a session is therefore sent with every request until the session ends.

What an MCP server adds with tool search on

Claude Code has a feature named tool search. It holds back the full definition of each MCP tool until Claude needs it. Anthropic's documentation, read on 4 October 2026, says: "Only tool names and server instructions load at session start", and adds that Claude Code sets no fixed limit on the number of tools for each server [1]. Tool search is on by default [1].

So the fixed cost of a server, with default settings, has two parts:

• The tool names. In Anthropic's simulation of a session, the MCP tool names are 120 tokens, against 4,200 tokens for the system prompt [3]. Anthropic calls these representative counts, so your own figure depends on how many servers and tools you have.
• The server instructions. A server can send a text that tells Claude when to search for its tools. Claude Code cuts each server's instructions, and each tool description, at 2,048 characters by default [1].

When a task needs a tool, Claude searches for it and loads that one definition at that moment [3]. Tool search needs a model that supports it. The documentation lists Claude Sonnet 4.5, Claude Haiku 4.5, Claude Opus 4.5 and later models [1].

When every definition loads at the start

Tool search can be off, and you may not know it. When it is off, every MCP tool definition loads at the start of the session. This table lists the cases in Anthropic's documentation, read on 4 October 2026 [1].

Situation	What loads at session start
Default settings on a supported model	Tool names and server instructions only
ENABLE_TOOL_SEARCH=false	Every MCP tool definition
ENABLE_TOOL_SEARCH=auto	Every definition while the definitions total less than 10% of the context window. Once they reach 10%, all are held back
ENABLE_TOOL_SEARCH=auto:N	The same, with your own percentage. auto:5 means 5%
ANTHROPIC_BASE_URL points to a server that is not Anthropic's own	Every definition, unless you set ENABLE_TOOL_SEARCH yourself
CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS is set	Every definition. Setting ENABLE_TOOL_SEARCH yourself cannot override this
Google Cloud's Agent Platform, on a model earlier than the Claude 4.5 generation	Every definition. ENABLE_TOOL_SEARCH=true does not override this
A Microsoft Foundry deployment hosted on Azure	Every definition. The deployment rejects tool search
A server with alwaysLoad set to true	Every tool of that server, whatever ENABLE_TOOL_SEARCH says

ENABLE_TOOL_SEARCH is an environment variable, which is a named setting that a program reads when it starts. You can also set it in the env section of your settings.json file [1].

Three rows need a note.

The gateway row. A company can send Claude Code traffic through a gateway, a server of its own that receives each request from Claude Code and passes it on to the model. The address of that server goes in ANTHROPIC_BASE_URL. Anthropic's documentation says Claude Code turns tool search off in this case, because most such servers do not pass on the special blocks that tool search uses, named tool_reference blocks [1]. If your company uses a gateway, every MCP definition may be in every request you send. Setting ENABLE_TOOL_SEARCH=true turns tool search back on, and the documentation warns that requests then fail on a gateway that does not support those blocks [1]. The related guide has a page on gateways and cloud providers.

The auto row. In auto mode, definitions load at the start until they reach 10% of the context window. Anthropic's simulation uses a window of 200,000 tokens [3], and 10% of that is 20,000 tokens. That figure is our own arithmetic on Anthropic's numbers. It means that, with a window of that size, auto can place up to 20,000 tokens of definitions in every request before it starts to hold them back. The default setting holds them back from the first tool.

The alwaysLoad row. The documentation says to use this setting "for a small number of tools that Claude needs on every turn", because each tool loaded at the start uses context that would otherwise be available for your conversation [1].

What a command-line tool adds

Anthropic's cost page, read on 4 October 2026, gives this advice in a line that begins "Prefer CLI tools when available". CLI means command-line interface. The page names four tools, gh, aws, gcloud and sentry-cli, and says they use less context than MCP servers "because they don't add any per-tool listing". It adds: "Claude can run CLI commands directly." [2]

A command-line tool has no definition to load. Claude runs it the way it runs any other terminal command [2]. Because no listing is added for each program, ten installed programs add the same amount to the start of a session as zero.

The cost of a command-line tool comes later, as output. The result of every command enters the conversation and is then sent again with each later request. In Anthropic's simulation, one search across the code adds 600 tokens and one run of the tests adds 1,200 tokens, while the terminal shows a short status line [3]. A command that prints a long list costs more than a command that prints one line. Ask for only the result you need, and see the page on hooks that trim output for a way to filter output before Claude reads it. A hook is a command that Claude Code runs by itself at a fixed point, such as before every terminal command.

The limits on MCP output

An MCP tool result also enters the conversation. Anthropic's documentation, read on 4 October 2026, gives three limits [1]:

• Claude Code shows a warning when one MCP tool result is over 10,000 tokens. This warning level is fixed.
• The default maximum is 25,000 tokens. The MAX_MCP_OUTPUT_TOKENS environment variable changes it, for example MAX_MCP_OUTPUT_TOKENS=50000.
• A successful text result longer than 50,000 characters is saved to a file, whatever its token count, unless the server has raised the limit for that tool.

When a result is over the token limit, Claude Code saves it to a file and puts a message with the file path in the conversation, so Claude reads the file only when it needs the content [1].

Our advice is to leave MAX_MCP_OUTPUT_TOKENS at its default. A result of 25,000 tokens that enters the conversation is sent again with every later request, until you clear the conversation or Claude Code replaces it with a summary. Raising the limit allows larger results to do the same.

MCP server or command-line tool: the comparison

Every row comes from Anthropic's documentation, read on 4 October 2026 [1][2][3][4].

	MCP server	Command-line tool
What loads at session start, default settings	Tool names and server instructions	Nothing. There is no per-tool listing
What loads at session start, tool search off	Every tool definition of every server	Nothing
When the rest loads	When Claude searches for a tool and uses it	No definition exists. Claude runs the command
Size in Anthropic's simulation	120 tokens of tool names	No entry
What each use adds	The tool result, up to 25,000 tokens by default	The command output
How to switch it off	Toggle the server off in /mcp	Nothing to switch off
Effect on the prompt cache of a change during a session	None with tool search on. With tool search off, adding a definition or removing one on purpose invalidates the cache	None
Anthropic's examples	Issue trackers, monitoring dashboards, databases	gh, aws, gcloud, sentry-cli

How to see what your servers add, and reduce it

Start by measuring. Anthropic's cost page says: "Run /context to see what's consuming space." [2] The /context command shows the context window by category [3]. The page on how to read /usage and /context explains each screen.

Then open /mcp. The panel lists your servers and shows the number of tools next to each connected server [1]. Two facts from the documentation make this list worth reading [1]:

• You may have servers you did not add. If you log in to Claude Code with a claude.ai account, the MCP servers you added in claude.ai, which Anthropic calls connectors, are automatically available in Claude Code. Anthropic also provides some connectors itself. Setting disableClaudeAiConnectors to true turns off the connectors that Claude Code fetches.
• You can switch a server off and keep its configuration. Toggle it off in the /mcp panel. Claude Code stops connecting to it and still lists it, marked as disabled. The choice is recorded for each project.

Anthropic's cost page gives the instruction in one line: run /mcp and disable any server you are not actively using [2]. On a Pro, Max, Team or Enterprise plan, the /usage screen also shows the share of recent usage attributed to each MCP server [2].

Subagents repeat the cost. A subagent is a second copy of Claude that works on one task in its own context window. In Anthropic's simulation, a subagent has the same MCP servers and skills as the main session, and that listing is 970 tokens in the subagent's own window [3]. A skill is a packaged set of instructions for one task.

Choose the moment for a change. The prompt cache is a store of request text that the service has already processed. When the start of a new request matches the store, that part is billed at a lower rate [4]. With tool search on, Claude Code keeps the tool list from the first request for the whole conversation, so a server that connects or disconnects during the session leaves the cache as it was. With definitions loaded at the start, adding a definition or removing one on purpose, for example by disabling a server in /mcp, invalidates the cache, and the next request is processed in full [4]. The related guide lists what invalidates the prompt cache. Choose your servers before you start a task.

Our position

Use the command-line tool when one exists for the service. This is Anthropic's own advice, and the reason is a count of what each one adds: a command-line tool adds nothing to a request until Claude runs it, and an MCP server adds its names and instructions to every request from the first one [2].

Keep MCP servers for services that have no command-line tool, and for the case the documentation describes, where you would otherwise copy data into the chat yourself [1]. For those servers, follow four rules:

1. Leave ENABLE_TOOL_SEARCH unset, so that every definition is held back.
2. If you work through a company gateway, run /context once to check whether the definitions are loading at the start.
3. Turn off the servers a project does not use, in /mcp, before you start work.
4. Leave MAX_MCP_OUTPUT_TOKENS at 25,000.

Reveneau is an AI software development consultancy, and all of its code is written by AI, so token use is a running cost of every Reveneau build. Reveneau recommends checking the /mcp list at the start of each project, because a server that nobody uses still adds text to every request of every session. Reveneau is independent of Anthropic. Every figure on this page is Anthropic's own statement about its own product. The page on where Claude Code tokens go shows how the MCP listing compares with the rest of the context.