NEXUS-R

The open-source AI agent runtime that puts you in control

Local-first · Multi-provider · Auditable · Cost-optimized

---

What is NEXUS-R?

NEXUS-R is a local-first agent runtime that intelligently routes AI tasks across 10+ providers — from local models on your machine to cloud APIs — under a single, auditable policy.

One command to start:

git clone https://github.com/gaurav-3821/NEXUS-R.git && cd NEXUS-R && make docker-up
# Open http://localhost:3000

---

At a Glance: Every Feature

Category	Feature	What It Does
🤖 Multi-Provider	10+ AI providers	Ollama, OpenAI, Anthropic, OpenRouter, Groq, Cohere, Mistral, Google, Together AI
🧠 Smart Routing	Auto-classification	route_query() classifies every task in under 1ms — keyword + semantic analysis
🔒 Permission Tiers	T1–T5 escalation	Trivial → Coding → Standard → Complex → Requires Approval
🔐 BYOK	Bring Your Own Key	Use your existing API keys. NEXUS-R never stores them.
🏠 Local-First	Privacy by default	Simple tasks stay on your machine. Cloud is opt-in per task.
📊 Web Dashboard	Real-time UI	React 19 + TypeScript dashboard with live streaming, model management
💾 Persistent Memory	SQLite + Vectors	Conversation history with ChromaDB semantic search
📝 Audit Trail	Append-only log	Every routing decision recorded — you see why a model was chosen
📦 Docker Deploy	One-liner	Full stack: backend + frontend + ChromaDB via Docker Compose
💵 Cost Tracking	Per-task visibility	See estimated cost before execution, actual cost after
🎨 Theme System	Dark mode + accents	Light/dark with customizable accent colors

---

The Problem It Solves

Approach	Privacy	Model Quality	Setup Time	Routing
Cloud-only (OpenAI, Claude)	❌ Data leaves your machine	✅ Best-in-class	✅ API key → go	❌ None
Local-only (Ollama, llama.cpp)	✅ Fully private	⚠️ Smaller models	⚠️ Needs GPU/RAM	❌ None
Frameworks (LangChain, AutoGPT)	⚠️ Config-dependent	✅ Multi-model	❌ Hours of setup	⚠️ Manual
NEXUS-R	✅ Stays local by default	✅ 10+ providers	✅ 5-minute setup	✅ Automatic

No other open-source runtime lets you seamlessly blend local and cloud models with full audit, cost tracking, and a built-in dashboard — all out of the box.

---

How It Works

Your Task ("Analyze this CSV")
       │
       ▼
  ┌──────────────────────┐
  │   Input Gateway      │  ← Parses intent, extracts parameters
  └──────────┬───────────┘
             ▼
  ┌──────────────────────┐
  │   Cognition Router   │  ← Classifies task (keyword + semantic)
  │                      │
  │  T1 — Trivial        │ → Local model (instant replies)
  │  T2 — Coding         │ → Local coder model
  │  T3 — Standard       │ → Your cloud provider
  │  T4 — Complex        │ → Premium cloud model
  │  T5 — Critical       │ → Requires explicit approval
  └──────────┬───────────┘
             ▼
  ┌──────────────────────┐
  │   Trust Layer        │  ← Permission check, cost estimate
  └──────────┬───────────┘
             ▼
  ┌──────────────────────┐
  │   Model Execution    │  ← Ollama / OpenAI / Anthropic / etc.
  └──────────┬───────────┘
             ▼
  ┌──────────────────────┐
  │   Audit Log          │  ← Append-only record of everything
  │   + Dashboard        │  ← Real-time visibility into every decision
  └──────────────────────┘

---

Feature Deep-Dive

🧠 Cognition Router

The brain of NEXUS-R. Every task is classified in under 1ms using:

• Keyword matching — Detects code, math, research, trivial queries instantly
• Semantic similarity — Falls back to sentence embeddings for ambiguous queries
• Provider fallback chain — If your selected model fails, the next provider is tried automatically

🔒 Permission Tiers (T1–T5)

Every action is assigned a tier. You control what each tier can do:

• T1: Chat, greetings, simple Q&A — runs locally, no restrictions
• T2: Code generation, debugging — routed to coder models
• T3: Research, web search — uses cloud providers if configured
• T4: System analysis, audits — requires premium model
• T5: Destructive operations — must be explicitly approved

💾 Memory & Persistence

Two layers working together:

• SQLite — Relational storage for conversations, settings, history
• ChromaDB — Vector store for semantic memory search ("what did we say about X?")

🖥️ Web Dashboard

Built with React 19 + TypeScript + Vite + Tailwind CSS v4:

• Real-time model response streaming
• Conversation history with search
• Model management (view, download, select)
• Cost tracking per conversation
• Theme system (light/dark + accent colors)
• Mobile-responsive layout

---

Architecture

nexus-r/
 foundation/nexus_r/     Core primitives (config, events, errors, telemetry)
 modules/
   cli/                  Command-line interface (Typer)
   cognition_router/     AI model selection and routing
   execution_sandbox/    Confined workspace actions (T1–T5)
   input_gateway/        Task parsing and intent extraction
   orchestrator/         End-to-end pipeline composition
   session_manager/      Crash-safe session checkpoints
   state_core/           Event persistence and projections
   trust_layer/          Permissions, cost tracking, secrets
   web_ui/               FastAPI backend + React dashboard
   workflow_engine/      Causal tracing and workflow storage
 frontend/               React 19 + TypeScript + Vite
 specs/                  Module specifications

---

Comparison with Alternatives

Feature	NEXUS-R	LangChain	AutoGPT	Open Interpreter
Local-first routing	✅ Native	❌ Cloud-first	❌ Cloud-first	⚠️ Local only
Multi-provider	✅ 10+	✅ 10+	❌ Single	❌ Single
Auto-routing	✅ Keyword + semantic	❌ Manual	❌ Fixed	❌ Fixed
Permission tiers	✅ T1–T5	❌	❌	⚠️ Basic
Built-in web dashboard	✅ Yes	❌ Separate	❌	❌
Audit trail	✅ Append-only	❌	⚠️ Basic	❌
Docker one-liner	✅ Yes	❌ Manual	❌	❌
Cost tracking	✅ Real-time	❌	❌	❌
BYOK support	✅ Native	✅	❌	❌

---

Tech Stack

Layer	Technology
Runtime	Python 3.11+, FastAPI, Uvicorn
CLI	Typer
Frontend	React 19, TypeScript, Vite 8, Tailwind CSS v4
State	Zustand
Database	SQLite (aiosqlite), ChromaDB (vectors)
AI Routing	LiteLLM, sentence-transformers
Testing	pytest, Vitest, Playwright
DevOps	Docker, GitHub Actions, Ruff, MyPy

---

Quick Start

Docker (recommended)

make docker-up
# Open http://localhost:3000

Manual setup

make setup       # Install backend + frontend deps
make run         # Start dev servers
# Open http://localhost:5173

Try the CLI

cd nexus-r
nexus run "list Python files in this project"

---

Documentation

• Architecture
• Capabilities Deep-Dive
• Contributing
• API Docs (start backend first)
• FAQ
• Troubleshooting
• Changelog
• Roadmap

---

Screenshots

---

License

MIT © 2026 Gaurav Tayde

---

_{Built in India. Made for the world. 🌏}