██████╗ ███╗   ███╗ ██████╗
    ██╔══██╗████╗ ████║██╔════╝
    ██████╔╝██╔████╔██║██║     
    ██╔═══╝ ██║╚██╔╝██║██║     
    ██║     ██║ ╚═╝ ██║╚██████╗
    ╚═╝     ╚═╝     ╚═╝ ╚═════╝
  

Predictive Minimal Context

Cut AI coding token costs by 40–96%. Drop-in proxy. Zero code changes.

The Problem · How It Works · Benchmarks · Quick Start · Integration · Research

Built by @mdayan24x

---

The Problem

"Uber deployed Claude Code to 5,000 engineers in December 2025. By April 2026 — just 4 months later — their entire annual AI budget was gone." — Forbes (source)

AI coding tools dump entire files into context. "Fix the race condition in login?" Here's 15 files — 45,000 tokens — when 500 would suffice. At scale: $500–$2,000/engineer/month. PMC fixes this structurally.

Approach	Problem
Prompt caching	Only helps reruns, not initial load
Vector embeddings	Lossy — misses call relationships
RAG chunking	Breaks logical boundaries
PMC (this project)	✅ AST-aware — 96% fewer tokens, same quality

---

How It Works

The 4 Tiers

Tier	Score	What the AI Gets	Example
T1 — Full Code	≥ 2.5	Complete function	def login(...) — 60 lines
T2 — Signature	≥ 1.0	name() → type	def verify(plain, hash) → bool
T3 — Stub	≥ 0.3	Name + location	[STUB] connect() → database.py:5
T4 — Omitted	< 0.3	Not sent	ConfigService, I18nService

Scoring: score = direct×3 + hop1×1.5 + hop2×0.6 + import×0.5 + config×1.0 + type×1.0 − cache×0.9

---

Benchmark Results

Codebase: FastAPI (48 files, 294 symbols, 33K LOC) · Model: DeepSeek V4 Flash

Naive = tokens Claude reads without PMC (files relevant to each task). PMC = compressed context.

Per task: 45K–148K (raw) vs 7.1K–7.8K (PMC)

45 requests: raw hits 3.2M, PMC stays at 130K

Task Details

Task	Files	Naive	PMC (measured)	Reduction
🟢 BackgroundTasks.add_task — validate None	4	45,000	7,119	84.2%
🟡 routing.py — fix nested function shadowing	7	85,000	7,836	90.8%
🔴 applications.py — middleware order validation	14	148,000	7,749	94.8%
Average	8.3	92,667	7,568	91.8%

Quality

100% score · 10/10 tasks passed · 36× cheaper ($1.82 → $0.05) · <5ms overhead

---

Quick Start

pip install pmc-engine

pmc index ./my-project                  # One-time index (~500ms)
pmc compress "fix the race condition"   # Compress a query
pmc serve --port 8080                   # Start proxy

# In another terminal:
export ANTHROPIC_BASE_URL="http://localhost:8080"
claude "fix the race condition in login"   # PMC compresses automatically

from pmc import PMCEngine
engine = PMCEngine()
engine.index("./my_project")
result = engine.compress("fix the race condition in login")
print(result.summary())  # 5,711 vs 45,000 naive (87.3%)

---

Integration

Tool	Method	Setup
Claude Code	HTTP proxy	export ANTHROPIC_BASE_URL="http://localhost:8080"
Claude Code	MCP server	Add pmc to mcpServers
Claude Code	Hooks	pmc install-cc-hooks
Cursor	MCP server	Same MCP config
Cline	apiBase	http://localhost:8080
Continue	apiBase	http://localhost:8080/v1
Aider	Env var	export ANTHROPIC_BASE_URL="http://localhost:8080"

---

Enterprise Savings

Company	Engineers	Without PMC	With PMC	Saved
Solo	1	$958/yr	$250/yr	$708
Startup	50	$48K/yr	$13K/yr	$35K
Scaleup	500	$479K/yr	$125K/yr	$354K
Enterprise	5,000	$9.85M/yr	$2.37M/yr	$7.49M

---

Research

Finding	Paper
20× compression, <1.5% loss	LLMLingua — EMNLP 2023 (arXiv)
11/13 models below 50% at 32K	NoLiMa — ICML 2025 (arXiv)
4× fewer tokens, +21.4% accuracy	LongLLMLingua — ACL 2024 (arXiv)
AST chunking beats naive	CAST — EMNLP 2025 (arXiv)
U-shaped attention in LLMs	Lost in the Middle — TACL 2024 (arXiv)

---

CLI

pmc index        # Build symbol index
pmc compress     # Compress a query
pmc serve        # Start HTTP proxy
pmc mcp          # Start MCP server
pmc bench        # Run benchmark
pmc verify       # Quality verification
pmc calibrate    # Auto-tune weights
pmc stats        # Show statistics

---

License

MIT — free. GitHub · PyPI · X/@mdayan24x

_{Built because AI coding costs are real, the problem is structural, and the fix is surgical.}