🌐 日本語
MCP Context Cost
IMPORTANT
→ Why: Context Rot mitigation (constant consumption of tool definitions pressures context) → Why: Knowledge Boundary mitigation (external knowledge retrieval reduces LLM's dependency on internal knowledge)
How MCP Consumes Context
When you connect an MCP server, tool definitions (name, parameter schema, description) are injected into the context window every turn. This is the same "resident cost" as CLAUDE.md.
| Property | Value |
|---|---|
| Injection Timing | Loaded as tool definitions at session start |
| Context Consumption | Constant consumption as tool definitions |
| How LLM Sees It | List of "available tools" |
| Danger Threshold | 20K+ tokens total across all MCPs |
Concrete Example of Context Consumption
WARNING
When MCP tool definitions balloon to 50K, the remaining budget drops to 93K — long conversations become impossible.
MCP as Knowledge Boundary Mitigation
While MCP carries Context Rot risk, it is also the most fundamental mitigation for Knowledge Boundary.
- Instead of relying on the LLM's internal knowledge (training data), directly reference external trusted sources
- Real-time access to API documentation, internal wikis, databases, etc.
- Address "unknown unknowns" by supplementing with external knowledge
Operational Best Practices
- Don't connect unnecessary MCP servers
- Monitor that total MCP consumption doesn't exceed 20K tokens
- Connect infrequently-used MCPs only when needed
Previous: Part 6: Context as Tool Definitions