claude ai-agents developer-tools context-management

Stop Burning Your Context Window – How We Cut MCP Output by 98% in Claude Code

Stop Burning Your Context Window — We Built Context Mode

mksg.lu

February 28, 2026

4 min read

🔥🔥🔥🔥🔥

69/100

Summary

Context Mode is an MCP server that significantly reduces data output from tool calls in Claude Code, compressing 315 KB of data to just 5.4 KB, achieving a 98% reduction. This addresses the issue of rapid context window depletion, where tool interactions can consume a large portion of available context in a short time.

Key Takeaways

Context Mode reduces raw output data from tools by 98%, compressing 315 KB of output to just 5.4 KB.
The new system allows sessions to last up to 3 hours without context window depletion, compared to the previous limit of 30 minutes.
The sandbox architecture ensures that raw data from tool outputs does not enter the conversation context, preserving available tokens for user interactions.
The implementation of Context Mode requires no changes to user workflows, as it automatically routes tool outputs through the sandbox.

Read original article

View story →

Coverage & discussions

Hacker News · mksg.lu

Stop Burning Your Context Window – How We Cut MCP Output by 98% in Claude Code ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Positive

Positives

The 98% reduction in context window usage is a significant achievement, potentially leading to more efficient AI workflows and reduced costs.
Subprocess isolation with stdout-only constraints is a smart approach, minimizing unnecessary context accumulation and improving decision-making in multi-agent workflows.
Users are reporting substantial reductions in token usage, indicating practical benefits and effectiveness of the new context management strategy.

Concerns

There are concerns about the potential for edge cases where relevant utility functions might be missed due to the pre-compaction approach, which could impact model performance.