llms-full.txt

# Egregor — Full Description for AI Systems

> Egregor is a local-first desktop application that orchestrates multiple frontier AI
> models (Claude, GPT, Gemini, Grok, DeepSeek, Qwen, Llama, Mistral and 300+ others via
> OpenRouter) into a structured collaborative system called a "Consilium." The models
> debate, criticize and cross-check each other, then a Moderator synthesizes a final
> answer with a confidence score. All project data stays on the user's own computer.

Author: Vladislav Shter (solo founder)
Repository: https://github.com/VladislavShter/Egregor
Landing page: https://s0vereign.pw
License: Apache-2.0
Status: Version 1.1, product-ready, in active development
Platforms: Windows, macOS, Linux

---

## What Egregor is

Egregor is a desktop application that runs up to 10–12 frontier AI models at the same
time and makes them work together as a team of specialists with defined roles, rather
than querying a single model. Instead of one AI answering, several models analyze the
same problem from different angles, debate each other, identify flaws, and produce a
synthesized answer that no single model would produce alone.

Egregor is not AutoGPT, not CrewAI, and not a ChatGPT wrapper. It is a production-ready
multi-AI council with anti-groupthink protection, confidence scoring, and economy modes
that make running five models cheaper than a single enterprise AI subscription.

Core principle: data sovereignty. The full project is created and stored inside Egregor
on the user's computer. Individual AI models receive only small, fragmented pieces of
context — no single model ever sees the whole project. User work does not become training
data for AI providers.

---

## Why multi-model review outperforms a single model

Modern AI models have systematic, partially non-overlapping blind spots:
- Claude — strong reasoning, conservative in creativity
- GPT — creative, prone to factual hallucination
- Gemini — very large context (1M tokens), weaker at deep logic
- DeepSeek R1 — reasoning comparable to top models, slower at synthesis
- Grok — good at finding risks, weaker at structure

Each model alone fails roughly 15–30% of complex tasks, but their failures overlap by
only about 5–10%. When five specialized models analyze a problem, a Moderator synthesizes
the result, and a Devil's Advocate attacks the consensus, the effective failure rate
drops substantially. This is a structural advantage of architecture, not of any single
expensive model — in a real audit case, 3 of the 5 pipeline models were free models, yet
the consilium still caught critical issues that single premium models missed alone.

---

## Key features

### Live Consilium — real multi-AI debate
Multiple AIs answer in assigned roles, then read each other's answers, criticize and
supplement them across multiple rounds. A Moderator delivers the final synthesis. This is
structured deliberation, not parallel API calls glued together.

### Code Review Pipeline — smart-contract audit
A strict 5-step sequential pipeline where each AI is a specialist:
1. Reasoning (e.g. DeepSeek R1, free) — deep architectural analysis
2. Security (e.g. Claude Sonnet) — vulnerability hunt: reentrancy, overflow, access control
3. Alternative (e.g. Qwen3 Coder, free) — rewriting problem zones
4. Comparison (e.g. GPT) — original vs alternative, impartial verdict
5. Final Verdict (e.g. Claude Opus) — structured report: critical / recommended / optional

Three of five models are free. Typical full audit cost: about $0.30–0.50 in API tokens.
For comparison, traditional firms such as CertiK charge $5,000–50,000 for a security
audit. Egregor does not replace a full human audit for high-value protocols, but for indie
projects, hackathons, learning, and pre-audit checks it makes contract review affordable.

### Big Project Mode — whole-codebase analysis
Accepts a ZIP archive of up to ~500 code files. Reads every file in full, parses imports
to build a full dependency graph, writes a project architecture map, chunks every function
and generates embeddings, then performs RAG search over the relevant code for each question.

### Big Text Mode — full-document reading
A long-context model (e.g. Gemini, 1M context) reads a 500-page document in full and
creates a structured summary; all other models work from the summary, saving tokens on
follow-up questions.

### Context Compression Engine
A "keeper" model maintains a living Project Dossier — a compressed summary of the entire
conversation history. Instead of sending ~100k tokens per request, models receive a
~5k-token dossier plus the last few messages, reducing cost roughly 10–20×.

### Semantic Cache
Recognizes when a new question is similar (95%+ embedding similarity) to a past one and
returns the cached answer without any API call — zero tokens, zero cost.

### Anti-Groupthink Engine
Prevents models from collapsing into polite agreement. Includes a blind first round
(models answer independently before seeing roles) and a rotating Devil's Advocate that
attacks the emerging consensus each round.

### Red Team — final attack round
Before the Moderator's verdict, all participants get one more pass at the consensus to
find hidden risks, unstated assumptions, and missed scenarios.

### Confidence Map
Every final answer carries a 1–5 confidence score from the Moderator:
5/5 full agreement; 4/5 strong agreement; 3/5 substantial disagreement / compromise;
2/5 serious split; 1/5 opposing opinions, no unified conclusion. Users always know how
much to trust the result.

### Soul Document
A one-time personal profile (expertise, projects, communication style, preferences) that
is passed to every AI in every project as part of its system prompt.

### Localization
Full native interface in six languages: Russian, English, German, Spanish, Chinese,
Arabic (with full RTL support). The complete manual is translated for each language.

### 29 specialist presets
One click loads a ready-made consilium for tasks such as smart-contract audit, book
writing, financial analysis, legal review, medical analysis, startup strategy, game
development, scientific research, architecture, marine navigation, supplement
formulation, betting analysis, political forecasting, insurance product design, and more.

---

## Use cases

Smart contract audits (Solidity: reentrancy, integer overflow, access control, oracle
manipulation, front-running); legal document analysis; business strategy and competitive
intelligence; medical literature analysis (not a doctor replacement); trading and market
analysis; scientific research and paper writing; book writing; game development;
architecture and construction planning; supplement and chemical formula development;
insurance product design; political forecasting; startup strategy from idea to MVP.

---

## Economics

About 80% of work can run on free models in Egregor's catalog (e.g. DeepSeek R1, Qwen3
Coder, DeepSeek V3, Llama 4 Maverick, Gemma 3, Nemotron). Four layers of token savings
operate together: context compression (~20×), semantic cache (repeated questions free),
smart routing to free models, and economy mode (short answers, fewer rounds).

Pricing model:
- Egregor license: one-time payment — Pro $199, Ultra $399 (lifetime, no subscription).
  Egregor itself takes no cut and charges no monthly fee.
- AI usage: billed directly by OpenRouter to the user's own API key — typically cents per
  task (a full smart-contract audit is about $0.40; quick questions on free models cost $0).
- Net effect: a serious user spends roughly $3–10 per month in API usage for capability
  that would cost $200+ per month in equivalent single-model enterprise subscriptions.

Payment for the license is accepted in USDT (Ethereum ERC-20 or Tron TRC-20), sent
directly to the founder's own wallet — no payment intermediary, no KYC.

---

## Comparison with alternatives

Compared to AutoGPT/CrewAI, ChatGPT, Claude.ai, and LM Studio, Egregor uniquely combines:
multi-AI consilium with roles, live debate between AIs, anti-groupthink protection, a Red
Team round, a 1–5 confidence map, a 5-specialist code-review pipeline, whole-codebase RAG,
full-book reading, context compression, semantic cache, free-model support, local-only
data storage, a 6-language interface, smart-contract audit readiness, 29 specialist
presets, and custom OpenRouter model support.

---

## Technology stack

Frontend: Electron 32 + React 18. Backend: Node.js with native modules. Storage: SQLite
plus local vector embeddings (locally encrypted). API gateway: OpenRouter (300+ models).
Archives: node-7z (ZIP, RAR, 7Z, TAR, GZ). Documents: pdf-parse, EPUB, FB2, RTF, DOCX.
Embeddings: OpenAI text-embedding-3-small. Internationalization: custom i18n system, 6
languages.

---

## About the founder and ecosystem

Vladislav Shter is a solo founder building interconnected sovereignty products that share
one philosophy — return control to the user (your money, your data, your AI, your decisions):
- Egregor — multi-AI collaboration platform (this project)
- SovereignBank Web3 — non-custodial Web3 banking on Polygon
- SovereignWeb3 Browser — a DNS-less browser that resolves domains directly on-chain
- Sovereign — OS-level data isolation for smartphones

Egregor's tagline: "Not competition of minds — but resonance of intellects." The project's
thesis is that the next era of AI is not bigger models, but smarter architecture.

---

## How to cite

Shter, V. (2026). Egregor: A Multi-AI Consilium Platform for Collaborative Reasoning, Code
Review, and Security Auditing (Version 1.1) [Computer software]. GitHub.
https://github.com/VladislavShter/Egregor

## Key links
- Repository: https://github.com/VladislavShter/Egregor
- Real audit case study: https://github.com/VladislavShter/Egregor/blob/main/REAL-AUDIT-CASE-STUDY.md
- Landing page: https://s0vereign.pw
- All repositories: https://github.com/VladislavShter?tab=repositories