New Full-SDLC Engineering Swarm & Incident Healing

The Sovereign Autonomous
AI Engineer & Agent System

Local-first LLM orchestration, air-gapped privacy, Sentry bug healing, and multi-agent swarm execution. Runs on LM Studio, Ollama, Claude 3.7, and GPT-4o.

$ pip install -e .
viki — reAct loop (Orythix Cortex)
viki> viki "fix sentry incident #409 in billing service"
[11:24:02] REFLEX Ingested stack trace from Sentry Webhook (NullPointer in invoice_calculator.py:L82)
[11:24:03] MCTS SWARM Spawning isolated Git Worktree: `worktree-fix-sentry-409`
[11:24:04] INCIDENT HEALER Executing AST Codemod & None guard injection
[11:24:06] PYTEST VERIFY Ran pytest tests/test_billing.py: 487 passed, 0 failed
✔ Resolution complete. Created PR #104 with green verification build.

Built for Production Agency

Designed from the ground up to replace generic chat wrappers with a deterministic, capability-aware sovereign AI engineer.

Air-Gapped Local Privacy

Default execution runs 100% on your hardware via LM Studio and Ollama. Enable VIKI_AIR_GAP=1 for strict zero outbound telemetry.

Multi-Model Routing

Intelligent provider fallback dynamically shifts simple queries to local models and complex reasoning tasks to Claude 3.7 or GPT-4o.

Autonomous Incident Healing

Ingests live Sentry & Datadog stack traces, reproduces errors inside clean Git worktrees, applies codemods, and verifies via pytest.

Neural Forge Prompt Baking

Bakes accumulated SQLite lessons and project wisdom directly into custom LM Studio model profiles without needing expensive GPU training.

Multi-Agent Swarms (MCTS)

Orchestrates leader-worker agent trees using Monte Carlo Tree Search for complex multi-file architectural refactoring tasks.

40+ Built-in Engineering Playbooks

Full-SDLC workflows covering spec-driven development, AST codemods, OpenAPI 3.1 / gRPC proto generation, and security auditing.

Supported LLM Inference Providers

Run completely offline with local open weights or scale out to cloud frontier APIs seamlessly.

Provider Supported Profiles / Models Privacy Mode Primary Recommendation
LM Studio gemma-4-e4b, qwen3.5-9b, deepseek-r1 100% Local / Air-Gapped Local primary default for daily coding & privacy
Ollama llama3, codellama, mistral, phi3 100% Local / Air-Gapped Zero-cost CLI automation & local background tasks
Anthropic claude-3-7-sonnet, claude-3-5-haiku Cloud API Complex multi-file refactoring & architecture
OpenAI gpt-4o, gpt-4o-mini, o3-mini Cloud API Deep technical review & TDD plan generation
Google Gemini gemini-2-5-pro, gemini-2-5-flash Cloud API Multimodal vision analysis & large context research

Deterministic Cognitive Data Flow

Every request passes through a 5-layer deliberation pipeline designed for maximum accuracy and zero unvetted actions.

LAYER 1
Perception
Ingests code, attachments, and user intent
LAYER 2
Judgment
Triages between Reflex & Deliberate paths
LAYER 3
Deliberation
Routes to optimal model & executes ReAct loop
LAYER 4
Reflection
Audits generated plans for logic & hallucination
LAYER 5
Meta-Cognition
Records lessons into SQLite RAG memory

How VIKI Compares

Architected specifically for sovereign developers who require complete ownership over their AI workflow.

Feature VIKI AutoGPT LangChain Manus / Devin
100% Air-Gapped Operation ✔ Yes ✖ No ✖ No ✖ No
LM Studio & Ollama Native ✔ Yes ✖ Partial ✖ Code only ✖ No
Automated Sentry Bug Healing ✔ Yes ✖ No ✖ No ✖ Proprietary
Neural Forge Prompt Baking ✔ Yes ✖ No ✖ No ✖ No
MCTS Multi-Agent Swarms ✔ Yes ✖ No ✖ Manual DAG ✖ Proprietary
Apache 2.0 Open Source ✔ Yes ✔ Yes ✔ Yes ✖ Closed

Get Started in 3 Steps

Launch VIKI locally on Windows, macOS, or Linux in less than two minutes.

1. Clone & Install

git clone https://github.com/Orythix/viki.git
cd viki
pip install -e .

2. Configure Models

cp .env.example .env
# Set LM Studio/Ollama host
# or add OpenAI/Claude keys

3. Launch CLI REPL

viki
# or run single query:
viki "fix bug in auth.py"