AI Briefing15% of Tokens Evaporate
Key point
Agents, skills, and MCP piled up, consuming 30,000 tokens before the session even started.
Details
Every Claude Code session loads agent, skill, MCP server, and memory file entirely into context, so a large cost was already being drained before any actual input.
- There were about 170 defined agents, but only around 10 had actually been used in the past week.
- Skills installed once and barely used, along with MCP servers not called in the last 7 days, were also still sitting there.
- The cumulative loss was about 30,000 tokens, roughly 15% of Sonnet's context window.
The problem is that this kind of waste is hard to see with typical usage tracking. ccusage shows consumption, but it doesn't reveal the fact that context is already being eaten into from the very start of each session.
The cleanup method was simple.
- Archive unused agents
- Disable MCP servers that aren't being called
In practice, the cleanup took just 5 minutes, and the key was directly checking 'what am I actually holding right now.'
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Foremerge v0.5.0 Released: Rust-Based Open-Source Tool Detects Intent Conflicts Between Parallel AI Coding Agents
Worktrunk CLI Released for Managing Git Worktrees for Parallel AI Coding AgentsClaude Code Update: Memory Warnings, MCP Timeout Controls, and Stability Improvements