llm-diet

Give Claude the right context upfront. Fewer turns, faster answers, lower cost.

Deterministic context retrieval for AI coding tools. Parses your repo into a call graph, intercepts every file read Claude makes, and returns compressed versions — so Claude explores freely but cheaply.

---

The Problem

Every Claude Code session starts blind. Claude explores your entire codebase before answering — reading files, listing directories, running commands. That exploration costs tokens and time.

Without llm-diet:
  Claude reads 10 files × 8,000 tokens = 80,000 tokens consumed
  Cost: $0.19 for a simple bug fix session

With llm-diet:
  Claude reads 10 files × 300 tokens  = 3,000 tokens consumed
  Cost: $0.025 for the same session

---

How It Works

User prompt
│
▼
context-engine (call graph)
│  scores every function against your query
▼
Claude Code session opens
│
▼
Claude calls read_file("validators/amazon.py")
│
▼
llm-diet-shadow MCP server intercepts
│  returns compressed 872-token version
│  instead of raw 6,590-token file
▼
Claude answers — correctly — using compressed context

Claude thinks it explored. It did — but every read returned our compressed version, not the raw file.

---

Benchmark

Tested on coupon-hunter-poc (40-node Python project)

File	Original	Compressed	Reduction
validators/playwright_amazon.py	6,590 chars	872 chars	86%
orchestrator.py	10,492 chars	2,169 chars	79%
connectors/playwright_amazon.py	3,067 chars	631 chars	79%
openrouter_agent.py	2,860 chars	966 chars	66%
retailmenot_scraper.py	2,705 chars	960 chars	64%
normalizer.py	1,074 chars	555 chars	48%

Overall: 32,856 → 10,044 chars across all indexed files 69% reduction — ~5,700 tokens saved per full codebase read

Session cost comparison (same task, same codebase — coupon-hunter-poc, verified live):

Task	diet-run	plain claude	Savings
Read file + explain function	$0.029	$0.033	12%
Read file + find bug + write fix	$0.071	$0.079	10%

Both sessions used the shadow MCP server for initial exploration. diet-run routes all file reads through compressed call-graph output — plain claude falls back to raw file reads when compressed output is truncated.

FastAPI benchmark (946-node project):

Query	Tokens without	Tokens injected	Reduction
fix error handling	946,210	221	>99.9%
add logging to pipeline	946,210	136	>99.9%

---

Quick Start

pip install llm-diet
cd your-project
mkdir .claude  # tells llm-diet Claude Code is present
context-engine install

Then open Claude Code normally:

claude

Or use enforced mode (blocks built-in Read tool, strict compression):

diet-run

---

What context-engine install does

1. Builds a call graph of your entire codebase (Python, JS, TS, JSX, TSX)
2. Writes .mcp.json — registers the shadow MCP server
3. Writes CLAUDE.md — tells Claude to use Low Bandwidth Mode

One command. No config files to edit manually.

---

The Shadow MCP Server

The shadow server is the core mechanism. It registers as an MCP server in .mcp.json. When Claude calls read_file, the shadow server intercepts it and returns:

• HIT (file is indexed): compressed call-graph version — signatures, first 8 lines of body, return/raise statements. Docstrings and comments stripped.
• MISS (file not indexed): raw file contents passed through unchanged
• Binary files: skipped with a one-line note
• Large unindexed files (>50k chars): truncated to 200 lines with an indexing suggestion

---

diet-run (enforced mode)

diet-run                    # current directory
diet-run /path/to/project   # specific directory

Launches Claude Code with:

• LLM_DIET_STRICT=1 — unindexed files return an error instead of raw content
• --disallowed-tools Read,Bash,Glob,Grep — filesystem exploration tools blocked
• Shadow server as the only file reader

Requires .cecl/graph.json and .mcp.json to exist (run context-engine install first).

---

Platform Support

Platform	Status
Claude Code	✅ Full support — shadow MCP server, diet-run, CLAUDE.md
Cursor	⚠️ Static rules file only (.cursor/rules/)
Windsurf	⚠️ Static rules file only (.windsurf/rules/)

---

Why Not RAG?

RAG requires embeddings, a vector database, and an LLM call in the retrieval path. llm-diet uses AST call graph analysis — deterministic, zero LLM calls at retrieval time, same query always returns the same result.

---

Commands

Command	What it does
context-engine install	Index codebase, write .mcp.json and CLAUDE.md
context-engine index .	Re-index after adding new files
context-engine watch .	Auto-reindex on file save
diet-run	Launch Claude Code in enforced Low Bandwidth Mode
diet-mcp	Start the shadow MCP server manually