entroly 0.8.5

The token saving proxy and context compression engine for AI coding agents. Reduce LLM API costs by 80% while providing full codebase context to Cursor, Claude Code, and Copilot.

🇨🇳 中文 • 🇯🇵 日本語 • 🇰🇷 한국어 • 🇧🇷 Português • 🇪🇸 Español • 🇩🇪 Deutsch • 🇫🇷 Français • 🇷🇺 Русский • 🇮🇳 हिन्दी • 🇹🇷 Türkçe

Entroly — Cut AI Token Costs by 70–95%

Your AI coding tools only see 5% of your codebase.
Entroly gives them the full picture — for a fraction of the cost.

npm install entroly-wasm && npx entroly-wasm  |  pip install entroly && entroly go  |  Live demo →

---

The Problem — and the Bottom-Line Impact

Every AI coding tool — Claude, Cursor, Copilot, Codex — has the same blind spot: it only sees 5–10 files at a time. The other 95% of your codebase is invisible. This causes hallucinated APIs, broken imports, missed dependencies, and wasted developer hours fixing AI-generated mistakes.

Models keep getting bigger — Claude Opus 4.7 just dropped with even more capability and even higher per-token costs. Larger context windows don't solve the problem; they make it worse. You're paying for 186,000 tokens per request — most of which is duplicated boilerplate.

Entroly fixes both problems in 30 seconds. It compresses your entire codebase into the AI context window at variable resolution, so your AI sees everything — and you pay for almost none of it.

---

What Changes on Day 1

Metric	Before Entroly	After Entroly
Files visible to AI	5–10	Your entire codebase
Tokens per request	~186,000	9,300 – 55,000
Monthly AI spend (at 1K req/day)	~$16,800	$840 – $5,040
AI answer accuracy	Incomplete, often hallucinated	Dependency-aware, correct
Developer time fixing AI mistakes	Hours/week	Near zero
Setup	Days of prompt engineering	30 seconds

ROI example: A 10-person team spending $15K/month on AI API calls saves $10K–$14K/month on day 1. Entroly pays for itself in the first hour. (It's free and open-source, so it actually pays for itself instantly.)

---

What Your Competitors Already Know

The teams adopting Entroly today aren't just saving money — they're compounding an advantage your team can't catch up to.

• Week 1: Their AI sees 100% of their codebase. Yours sees 5%. They ship faster.
• Month 1: Their runtime has learned their codebase patterns. Yours is still hallucinating imports.
• Month 3: Their installation is plugged into the federation — absorbing optimization strategies from thousands of other teams worldwide. Yours doesn't know this exists.
• Month 6: They've saved $80K+ in API costs. That budget went into hiring. You're still explaining to finance why the AI bill keeps growing.

Every day you wait, the gap widens. The federation effect means early adopters get smarter faster — and that advantage compounds.

---

How It Works (30 Seconds)

npm install entroly-wasm && npx entroly-wasm
# or
pip install entroly && entroly go

That's it. Entroly auto-detects your IDE, connects to Claude/Cursor/Copilot/Codex/MiniMax, and starts optimizing. No configuration. No YAML. No embeddings.

What happens under the hood:

1. Index — Maps your entire codebase in <2 seconds
2. Score — Ranks every file by information density
3. Select — Picks the mathematically optimal subset for your token budget
4. Deliver — Critical files go in full, supporting files as signatures, everything else as references
5. Learn — Tracks what works, gets smarter over time

Your AI now sees 100% of your codebase. You pay for 5–30% of the tokens.

---

The Competitive Edge — What Sets Entroly Apart

🧠 It Gets Smarter Without Costing You More

Most "self-improving" AI tools burn tokens to learn — your bill grows with their intelligence. Entroly's learning loop is provably token-negative: it cannot spend more on learning than it saves you.

The math is simple and auditable:

Learning budget ≤ 5% × Lifetime savings

Day 1: 70% token savings. Day 30: 85%+. Day 90: 90%+. The improvement costs you $0.

🌐 Federated Swarm Learning — The Part That Sounds Like Science Fiction

Now take the Dreaming Loop and multiply it by every developer on Earth who runs Entroly.

While you sleep, your daemon dreams — and so do 10,000 others. Each one discovers slightly different tricks for compressing code. Each one shares what it learned — anonymously, privately, no code ever leaves your machine. Each one absorbs what the others found.

You wake up. Your AI is smarter than when you left it. Not because of anything you did — because of what the swarm dreamed.

Your daemon dreams → discovers a better strategy → shares it (anonymously)
     ↓
10,000 other daemons did the same thing last night
     ↓
You open your laptop → your AI already absorbed all of it

Network effect:

• Every new user makes everyone else's AI better — that installed base can't be forked
• Your code never moves. Only optimization weights — noise-protected and anonymous
• Infrastructure cost: $0. It runs on GitHub. No servers. No GPUs. No cloud

# Opt-in — your choice, always
export ENTROLY_FEDERATION=1

✂️ Response Distillation — Save Tokens on Output Too

LLM responses contain ~40% filler — "Sure, I'd be happy to help!", hedging, meta-commentary. Entroly strips it. Code blocks are never touched.

Before: "Sure! I'd be happy to help. Let me take a look at your code.
         The issue is in the auth module. Hope this helps!"

After:  "The issue is in the auth module."
         → 70% fewer output tokens

Three intensity levels: lite → full → ultra. Enable with one env var.

🔒 Runs Locally. Your Code Never Leaves Your Machine.

Zero cloud dependencies. Zero data exfiltration risk. Everything runs on your CPU in <10ms. Works in air-gapped and regulated environments — nothing ever phones home.

---

Works With Your Stack

Tool	Setup
Cursor	entroly init → MCP server
Claude Code	claude mcp add entroly -- entroly
GitHub Copilot	entroly init → MCP server
Codex CLI	entroly wrap codex
Windsurf / Cline / Cody	entroly init
Any LLM API	entroly proxy → HTTP proxy on localhost:9377

Also: OpenAI API • Anthropic API • LangChain • LlamaIndex • MCP-native

---

Benchmarks

Live Evolution Trace

This is from this repo's vault, not a roadmap:

[detect]     gap observed → entity="auth", miss_count=3
[synthesize] StructuralSynthesizer ($0, deterministic, no LLM)
[benchmark]  skill=ddb2e2969bb0 → fitness 1.0 (1 pass / 0 fail, 338 ms)
[promote]    status: draft → promoted
[spend]      $0.0000 — invariant C_spent ≤ τ·S(t) holds

Accuracy Retention

Compression doesn't hurt accuracy — we measured it (n=100, gpt-4o-mini, Wilson 95% CIs):

Benchmark	Baseline (95% CI)	With Entroly (95% CI)	Retention
NeedleInAHaystack	100% [83.9–100%]	100% [83.9–100%]	100.0%
GSM8K	85.0% [76.7–90.7%]	86.0% [77.9–91.5%]	101.2%
SQuAD 2.0	84.0% [75.6–89.9%]	83.0% [74.5–89.1%]	98.8%

Confidence intervals overlap on every benchmark — accuracy is statistically indistinguishable from baseline. Reproduce: python -m bench.accuracy --benchmark all --model gpt-4o-mini --samples 100

CI/CD Integration

Run token cost checks in every PR — catch regressions before they ship:

- uses: juyterman1000/entroly-cost-check-@v1

→ entroly-cost-check GitHub Action

---

Watch It Run — Live Notifications

Three chat integrations ship in the box. See every gap detection, skill synthesis, and dream-cycle win in real-time:

export ENTROLY_TG_TOKEN=...          # Telegram (2-way: /status /skills /gaps /dream)
export ENTROLY_DISCORD_WEBHOOK=...   # Discord
export ENTROLY_SLACK_WEBHOOK=...     # Slack

---

Portable Skills (agentskills.io)

Skills Entroly creates aren't locked in. Export to the open agentskills.io v0.1 spec:

node node_modules/entroly-wasm/js/agentskills_export.js ./dist/agentskills
python -m entroly.integrations.agentskills ./dist/agentskills

Every exported skill carries origin.token_cost: 0.0 — the zero-cost provenance travels with it.

---

Full Parity: Python & Node.js

Both runtimes are feature-complete. Same engine, same vault, same learning loop:

Capability	Python	Node.js (WASM)
Context compression	✅	✅
Self-evolution	✅	✅
Dreaming loop	✅	✅
Federation	✅	✅
Response distillation	✅	✅
Chat gateways	✅	✅
agentskills.io export	✅	✅

---

Deep Dive

Architecture, 21 Rust modules, 3-resolution compression, provenance guarantees, RAG comparison, full CLI reference, Python SDK, LangChain integration → docs/DETAILS.md

---

Stop paying for tokens your AI wastes. Start running an AI that teaches itself.
npm install entroly-wasm && npx entroly-wasm  |  pip install entroly && entroly go

Discussions • Issues • Apache-2.0 License