This 98% Token Cut Tool Exposes Claude Code's Dirty Secret
I've watched Claude Code sessions die a slow, expensive death more times than I care to count. You know the drill: start with a crisp 200,000-token context window, watch it fill with garbage from Model Context Protocol (MCP) servers, then suffer through degraded performance as Claude forgets your instructions and makes rookie mistakes.
Turns out someone finally got fed up enough to do something about it.
Developer mksg just dropped a tool called Context Mode that cuts MCP output pollution by 98%. Not 50%. Not 80%. Ninety-eight percent. The project has already racked up 228 GitHub stars since launch, which tells you how desperate developers are for this fix.
The MCP Money Pit
Here's what nobody talks about: MCP tool calls don't just use tokens—they hemorrhage them. We're talking tens of thousands of tokens consumed by server descriptions alone, before you even start coding. That's like showing up to a 200k-word essay contest and spending half your allowance introducing yourself.
Context Mode solves this with a brutally simple approach: isolated subprocesses where only stdout enters the context window. Everything else? Gone.
<> "Subagent routing shows massive differences in output reduction, with real-world data showing auto-upgrading Bash subagents to use batch_execute as a key factor."/>
The subagent routing piece is particularly clever. Instead of dumping raw tool output into Claude's working memory, the system spins up isolated agents that handle the grunt work and only report back the essentials.
Anthropic's Band-Aid Approach
This isn't the first time Claude Code's context management has needed emergency surgery. Earlier in 2025, Anthropic rolled out:
- Context editing: 84% token reduction, 29-39% performance boost
- Plan mode: Halves token usage
- PreCompact hooks: 40% productivity improvement
But here's my issue: these are all reactive fixes. Why is a 200,000-token context window filling up so fast that we need aggressive compaction at 75% utilization instead of the old 90%+ threshold?
The Skills vs. MCP War
Smart developers are already jumping ship from MCP to Skills—Anthropic's more efficient alternative that uses just 30-50 tokens for metadata until actually triggered. Skills can load dozens of tools without bloat, while MCP servers stuff your context like a Thanksgiving turkey.
The community has developed workarounds that feel more like emergency medicine:
/compactcommands that recover 70% of context- Manual
.claudeignorefiles to cut 25% of token waste - Constant
/contextmonitoring like checking your bank balance
These shouldn't be necessary.
What This Really Means
Context Mode's success exposes a fundamental design flaw in Claude Code's architecture. When a solo developer can achieve 98% reduction in tool output pollution, it suggests Anthropic either didn't prioritize this problem or didn't understand its scope.
The 159 points and 43 comments on Hacker News show this isn't just a niche developer problem. People are genuinely excited about basic context hygiene.
Users report that aggressive pruning "restores output quality," which is a polite way of saying Claude Code gets stupid when context fills up. The fact that we need a 50k token reasoning buffer in a 200k window is telling.
My Bet
Anthropic will quietly integrate Context Mode's approach into Claude Code within six months, probably without crediting mksg. The 228 GitHub stars and developer enthusiasm create pressure they can't ignore. Meanwhile, MCP will get a quiet redesign to be less context-hungry—but not before more indie developers build their own solutions to Anthropic's architectural choices.
