The 4-tool stack that cuts Claude Code context waste

The four leaks 🕳

And it is not just cost. A crowded context makes answers drift too, so plugging these leaks keeps Claude sharp deeper into long sessions.

The stack at a glance 📚

They are fully independent, so install one or all four. Below is the order I would do it in, quickest win first. All four together take about ten minutes.

1. caveman: short answers 🪨

caveman is a Claude Code plugin that makes Claude answer in tight caveman-speak while keeping your code, commands, and error messages byte-for-byte exact. It fires on every reply, from message one, no prompt changes needed. Their own tagline: "why use many token when few do trick".

Run this in your terminal:

claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman

Check it works: ask Claude anything and the reply comes back short. Say "normal mode" to switch it off, type /caveman to bring it back.

2. context-engineering: lean sessions 🧩

context-engineering is a plugin of 17 tiny skills that stay asleep until your task matches one, so the right skill loads and the rest stay out of Claude's head. It fires automatically: Claude activates a skill only when the task calls for it, and nothing loads up front.

Run these inside Claude Code (they are slash commands):

/plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering

/plugin install context-engineering@context-engineering-marketplace

Check it works: type /plugin and context-engineering should sit under your installed plugins.

3. symdex: pinpoint lookups 🎯

symdex indexes your repo into one small local file (a map of every function, class, and route) so Claude jumps straight to the exact code it needs instead of reading whole files to get oriented. It fires whenever Claude would normally open files to hunt for something: it queries the map instead.

Run this in your terminal, inside your project folder:

claude mcp add symdex -- uvx symdex serve

Check it works: type /mcp inside Claude Code and symdex should show as connected. Then say "index this project with symdex" once. After that, every search returns its own token-savings estimate, so you see real numbers instead of trusting anyone's headline.

4. lean-ctx: cheap re-reads 🔁

lean-ctx runs a small local layer between Claude and your files that caches everything Claude reads, so the first read costs full price and every reopen is nearly free (about 13 tokens, their number). It fires automatically on every file read and shell output, across sessions. I put it last because it is the deepest install of the four (shell hooks plus a little background daemon), so give it five calm minutes.

Run this in your terminal:

curl -fsSL https://leanctx.com/install.sh | sh
lean-ctx wrap claude

Prefer Homebrew? brew tap yvgude/lean-ctx && brew install lean-ctx replaces the first line. The wrap command registers the MCP server, hooks, and daemon in one go, and lean-ctx unwrap claude undoes the whole thing.

Check it works: after your next Claude Code session, run lean-ctx gain in the terminal. It shows tokens and dollars saved in real time.

Honest numbers 🧮

• 🔁 lean-ctx claims 60 to 90% fewer tokens on reads and shell output, with the receipts in the benchmarks section of their README
• 🪨 caveman benchmarks 65% fewer output tokens on average (range 22 to 87%), and they openly warn the skill itself adds about 1 to 1.5k input tokens per turn, so whole-session savings run smaller. Their honest numbers doc is genuinely refreshing
• 🎯 symdex skips headline claims and prints a savings estimate after each search, so your numbers are your own
• 📉 The "bill dropped 10x" story is the original dev's result for his own workflow. Treat it as a ceiling, not a promise

The links 🔗

🔁 lean-ctx: github.com/yvgude/lean-ctx
🪨 caveman: github.com/JuliusBrussee/caveman
🎯 symdex: github.com/husnainpk/SymDex
🧩 context-engineering: github.com/muratcankoylan/Agent-Skills-for-Context-Engineering
💡 The original stack idea: @speedy_devv