Connecting more tools can make an agent less effective: their descriptions consume context that could otherwise hold code and conversation. A tool description defines an operation the agent can invoke. In Model Context Protocol (MCP), tool definitions are JSON schemas describing what a tool does, its parameters, and its return value. section 9.5.1 section 3.3.5
MCP tool definitions can load into the model’s context window at session start, whether or not the agent uses every tool. A server with twenty tools is estimated to consume roughly 10,000–15,000 tokens; five such servers could consume 50,000–75,000. Less space for code and conversation can lead to earlier compaction and poorer results on complex tasks. section 9.5.1
A crowded coding session
A backend team connected GitHub, PostgreSQL, internal documentation, Sentry, and Jira MCP servers. The /context command in Claude Code showed that MCP tools occupied 58,400 tokens, or 29% of the context window. The team had less room for code and conversation, and sessions became forgetful sooner. section 9.5.1
Table 9.6 Context Usage claude-sonnet-4-5
|
Source
|
Tokens
|
Share
|
|
System prompt
|
3,900 tokens
|
2%
|
|
MCP tools
|
58,400 tokens
|
29%
|
|
Memory files
|
1,200 tokens
|
1%
|
|
Messages
|
42,000 tokens
|
21%
|
|
Free space
|
94,500 tokens
|
47%
|
Compare the tool-definition cost with messages, system prompt, memory files, and remaining space in the measured session. section 9.5.1
Reduce the cost and risk of exposed capabilities with a few deliberate choices:
- Prefer CLI tools when they provide the same functionality without persistent MCP schema overhead. section 9.5.1
- Choose a small set of integrations and vet MCP servers like production dependencies; every connection is also a potential attack surface. section 3.3.5
- Use an MCP proxy to present a single interface and dynamically load relevant capabilities, rather than exposing every tool at once. section 3.3.5
- Use an MCP gateway where centralized authentication, access policies, audit logging, and suspicious-activity blocking are needed. section 3.3.5



