token-stack.md

The Token-Saving Stack

Three tools that work together to cut your Claude Code token usage by 60-90%. Install all three once. They run automatically after that.

---

1. RTK: Rust Token Killer

Intercepts every Bash command before it hits your context window and compresses the output. Transparent: you run git status, RTK runs it and strips the noise.

Install:

# Mac/Linux
curl -fsSL https://rtk.ai/install.sh | sh

# Windows (PowerShell)
irm https://rtk.ai/install.ps1 | iex

Wire into Claude Code:

rtk init -g

Verify:

rtk init --show    # all items should show [ok]
rtk gain           # shows token savings analytics

---

2. jMunch Family: Smart File Reading

Replaces Bash cat/grep/find with token-optimized readers. Each tool understands its file type and returns only what Claude needs.

Tool	For	Token savings
jCodeMunch	.py .ts .tsx .js	50,000+ tokens per call
jDocMunch	.md .mdx .rst	Sections only, not full files
jDataMunch	.json .html .csv	Schema + sample, not full dump

Install:

# Via Claude Code MCP config (~/.claude/claude_desktop_config.json)
# Add the jcodemunch-mcp server: see: https://github.com/jgravelle/jcodemunch-mcp

Use in CLAUDE.md:

## Tool Priority
1. Code files → jCodeMunch (mcp__jcodemunch__*)
2. Doc files → jDocMunch (mcp__jdocmunch__*)
3. Data files → jDataMunch (mcp__jdatamunch__*)

---

3. fff: Fast File Finder

Frecency-ranked file search (frequent + recent files first). Replaces find and ls -r with a single MCP call. Git-dirty files boosted to top.

Install:

# Download binary from: https://github.com/dmtrKovalenko/fff.nvim
# Add fff-mcp to your MCP config

Use in CLAUDE.md:

## File Search
- Find files by name → fff (mcp__fff__find_files)
- Search file contents → fff grep (mcp__fff__grep)
- NEVER use Bash grep or find

---

Combined CLAUDE.md Block

Copy this into your project's CLAUDE.md to wire all three at once:

## Token-Efficient Navigation

### Tool Priority
1. Code files (.py/.ts/.tsx) → jCodeMunch (`mcp__jcodemunch__*`)
2. Doc files (.md/.mdx/.rst) → jDocMunch (`mcp__jdocmunch__*`)
3. Data files (.json/.html) → jDataMunch (`mcp__jdatamunch__*`)
4. File search by name → fff (`mcp__fff__find_files`)
5. File content search → fff grep (`mcp__fff__grep`)

### Bash: allowed ONLY for
- git commands
- Package installs (npm install, pip install)
- CLI execution and build runners

### RTK hook active
All Bash output is compressed by RTK before hitting context.
Run `rtk gain` to see savings.

---

Built by Albatross AI: tools for contractors and builders who don't want to burn money on tokens.