--- title: "Unused MCP servers" description: "antiburn flags sessions that loaded an MCP server's tools into context but never called any of them." url: "https://antiburn.ai/docs/findings/unused-mcp-servers/" --- ## What it detects Connecting an MCP server to your agent loads that server's tool definitions into every session's context, whether or not the session ends up using them. antiburn checks, for each session with real assistant work in it, whether every loaded MCP server was actually invoked at least once. A session gets flagged as soon as one loaded server sits unused for the whole session. This is an absence claim - antiburn is asserting a server was never called - so it only fires from complete evidence. A session where antiburn only partially read the tool-invocation or context-loading records never produces this finding, even if what it did read looks unused, because a missed record could be the one call that used it. ## Why it burns tokens An MCP server's tool definitions occupy context space in every request for the life of the session, whether you call the server once, many times, or never. A server loaded and never invoked is pure overhead - context you paid to send on every turn for nothing. ## How to fix it Only connect the MCP servers a given project or session actually needs, and disconnect the rest. Most agents let you scope MCP servers per project or disable them per session rather than loading every configured server everywhere by default - check your agent's own MCP configuration for how it scopes servers, since this varies by agent. ## When it shows "not assessed" Today this only assesses on agents whose logs attribute a tool call back to the MCP server that loaded it - among antiburn's supported agents, that's Claude Code. Agents that don't capture this attribution can't be assessed for this finding at all. Even on a supported agent, a session shows not assessed rather than clean when antiburn's read of the session is incomplete - a partial tool-invocation log or a partial record of which servers were loaded - because an absence claim needs complete evidence to be trustworthy. Sessions with no real assistant work in them (idle sessions) aren't counted either way. See [hygiene score and findings](/docs/concepts/hygiene-score-and-findings/) for how this fits into your overall score.
Unused MCP servers
antiburn flags sessions that loaded an MCP server's tools into context but never called any of them.
What it detects
Connecting an MCP server to your agent loads that server’s tool definitions into every session’s context, whether or not the session ends up using them. antiburn checks, for each session with real assistant work in it, whether every loaded MCP server was actually invoked at least once. A session gets flagged as soon as one loaded server sits unused for the whole session.
This is an absence claim - antiburn is asserting a server was never called - so it only fires from complete evidence. A session where antiburn only partially read the tool-invocation or context-loading records never produces this finding, even if what it did read looks unused, because a missed record could be the one call that used it.
Why it burns tokens
An MCP server’s tool definitions occupy context space in every request for the life of the session, whether you call the server once, many times, or never. A server loaded and never invoked is pure overhead - context you paid to send on every turn for nothing.
How to fix it
Only connect the MCP servers a given project or session actually needs, and disconnect the rest. Most agents let you scope MCP servers per project or disable them per session rather than loading every configured server everywhere by default - check your agent’s own MCP configuration for how it scopes servers, since this varies by agent.
When it shows “not assessed”
Today this only assesses on agents whose logs attribute a tool call back to the MCP server that loaded it - among antiburn’s supported agents, that’s Claude Code. Agents that don’t capture this attribution can’t be assessed for this finding at all.
Even on a supported agent, a session shows not assessed rather than clean when antiburn’s read of the session is incomplete - a partial tool-invocation log or a partial record of which servers were loaded - because an absence claim needs complete evidence to be trustworthy. Sessions with no real assistant work in them (idle sessions) aren’t counted either way.
See hygiene score and findings for how this fits into your overall score.