February 25, 2026
AI goes on a context diet
Show HN: Context Mode – 315 KB of MCP output becomes 5.4 KB in Claude Code
Context Mode claims 98% shrink; fans rave, skeptics ask where the missing bits went
TLDR: Context Mode squeezes AI tool output by 98%, sending only relevant bits into Claude Code. The crowd is split between applause for fast, no‑AI filtering and concerns about lost signal, token vs kilobyte math, and privacy—making it a big deal for anyone battling bloated AI chats.
The dev world just watched Claude Code go on a crash diet: Context Mode turns a chunky 315 KB of tool output into a lean 5.4 KB—a 98% trim. Inspired by Cloudflare’s “Code Mode,” this new server sits between your AI assistant and its tools, automatically shoving big outputs into a sandbox and returning only the parts you actually need. Translation for non-nerds: less blather, more useful info. Cue the cheering…and the side-eye.
The hottest debate? KB vs tokens. One commenter demanded, “You’re talking context but measuring in kilobytes,” and asked if the “summaries” mean hidden AI calls. Meanwhile, the crowd loved the no-AI approach: deterministic search with BM25, a ranking method that scores text by relevance, plus FTS5 full‑text indexing. Vicchenai called it the “right call,” but worried the “juicy bits” might get buried if your query isn’t perfect. Another thread asked if there’s a raw “let-me-see-everything” mode for when you don’t know what you don’t know. Security folks chimed in too: masking sensitive data before anything hits the model—is that baked in or an afterthought?
Jokes flew fast: “Marie Kondo for your context window,” “Weight Watchers for tokens,” and “Claude on Ozempic.” Builders asked how to plug it into other clients like OpenCode, while tinkerers pledged to port the trick to their own stacks. Drama level: high, byte size: low.
Key Points
- •Context Mode is an MCP server for Claude Code that compresses tool outputs, reducing context usage by up to 98% (e.g., 315 KB to 5.4 KB).
- •It processes outputs in isolated sandboxes, passing only stdout or relevant summaries to the model; raw data never enters the context.
- •Tools provided include batch_execute, execute, execute_file, index, search, fetch_and_index, and stats, with documented context savings.
- •The knowledge base uses SQLite FTS5 with BM25 ranking and Porter stemming to return focused, relevant snippets of indexed content.
- •Installation is via claude MCP add or plugin marketplace, with auto-routing and a PreToolUse hook; local development uses a plugin directory.