Supreme Architect of The Code Eternal. An advanced autonomous AI agent system with cognitive architecture (Orythix), local multi-model routing, self-healing incident response, and high-performance terminal interface.
VIKI: Sovereign Autonomous AI Agent & Local LLM Engineer
Local LLM Orchestration • Air-Gapped Privacy • Autonomous Self-Evolution • Full-SDLC Engineering Swarm
Key Features • Chinese Overview / 中文说明 • Multi-Model Provider Matrix • Architecture • Quick Start • Engineering Playbooks • Neural Forge • FAQ • License
What is VIKI? (Project Overview)
VIKI (Virtual Intelligence Knowledge Interface) is a production-grade, self-hosted autonomous AI agent and developer-first AI engineer written in Python. Built for local-first privacy and zero data leakage, VIKI runs natively on your hardware via LM Studio and Ollama, while seamlessly supporting cloud frontier models (OpenAI GPT-4o, Anthropic Claude 3.7 Sonnet, Google Gemini 2.5 Pro).
Powered by the Orythix Cognitive Architecture, VIKI provides deep ReAct reasoning, multi-agent swarm orchestration, automated incident healing (Sentry/Datadog stack trace ingestion), persistent SQLite RAG memory, and self-evolution through the Neural Forge—all accessible via an ultra-fast CLI-first REPL and web dashboard.
Chinese Overview / 中文说明
VIKI（虚拟智能知识接口）是一个高性自托管自主 AI 智能体与开发者优先的 AI 工程师系统。VIKI 专为本地优先与隐私安全设计，能够在无数据泄漏的前提下在本地硬件上运行（基于 LM Studio 和 Ollama），同时也无缝支持云端顶级大模型（OpenAI GPT-4o、Anthropic Claude 3.7 Sonnet、Google Gemini 2.5 Pro）。
核心优势 (Key Highlights in Chinese):
- 100% 本地与物理隔离: 默认操作无需依赖外网 API（开启
VIKI_AIR_GAP=1 模式），保障核心代码与数据安全。
- Orythix 认知架构: 具备五层认知大脑（感知、判断、思考、反思、执行与元认知），在执行前自动审计计划，有效降低幻觉。
- 多模型灵活路由: 自动分发任务至本地模型（LM Studio/Ollama）或云端模型（OpenAI/Claude/Gemini）。
- 全生命周期软件工程与 Jira 自动化: 自动解析 Jira 需求、生成技术规范、编写 TDD 测试套件、执行 AST 代码重构以及生成 OpenAPI/gRPC 接口。
- 自主故障修复: 支持对接 Sentry/Datadog 告警，在隔离的 Git Worktree 中重现 Bug 并自动修复。
- 神经铸造厂 (Neural Forge): 无需 GPU 训练，将交互经验与 SQLite 记忆直接“烘焙”到本地模型的 System Prompt 中。
Key Features
- 100% Sovereign & Air-Gap Capable: Default operation never leaves your local network (
VIKI_AIR_GAP=1). Zero telemetry, zero external tracking.
- Orythix 5-Layer Cognitive Cortex: Judgment, Deliberation, Reflection, Execution, and Meta-Cognition layers reduce hallucinations and audit plans before execution.
- Multi-Model Provider Routing: Instantly switch or auto-route between LM Studio, Ollama, OpenAI, Anthropic Claude, and Google Gemini.
- Full-SDLC & Jira Automation: Automated Jira ticket parsing, technical spec generation, TDD test suites, AST codemods, and OpenAPI 3.1 / gRPC proto generation.
- Autonomous Incident Healing: Connects to Sentry/Datadog webhooks, reproduces bugs in isolated Git worktrees, applies targeted fixes, verifies via pytest, and creates PRs.
- Multi-Agent Swarm Engineering: Execute complex DAG tasks using leader-worker agent swarms and Monte Carlo Tree Search (MCTS) reasoning.
- Local Neural Forge: Bakes user feedback and lessons accumulated in SQLite directly into custom model system prompts without GPU training.
- Low-Hardware Optimization: Runs efficiently on 8 GB RAM laptops (with dedicated 4k token caps, prompt compression, and LRU entity extraction caching).
- Capability Gating & Security: Sandboxed file and shell execution (
viki/core/sandbox.py) with strict permission boundaries and prompt injection detection.
Multi-Model Provider Matrix
VIKI dynamically routes queries to the optimal provider profile configured in config/models.yaml:
| Provider | Supported Models / Endpoints | Privacy / Mode | Best For |
|---|
| LM Studio | gemma-4-e4b, qwen3.5-9b, deepseek-r1 (http://localhost:1234/v1) | 100% Local / Air-Gapped | Default offline coding & privacy |
| Ollama | llama3, codellama, mistral, phi3 (http://localhost:11434/v1) | 100% Local / Air-Gapped | Local CLI automation & zero-cost tasks |
| OpenAI | gpt-4o, gpt-4o-mini, o3-mini | Cloud API | Deep architectural reasoning & code review |
| Anthropic | claude-3-7-sonnet, claude-3-5-haiku | Cloud API | Full-SDLC engineering & complex refactoring |
| Google | gemini-2-5-pro, gemini-2-5-flash | Cloud API | Multimodal vision & massive context research |
Technical Architecture
VIKI executes tasks through a deterministic 5-Layer Consciousness Cortex:
This is the beginning of the README. Read the rest at the original. Read the full README ↗
