nexus-ai-fs 0.5.2

AI-Native Distributed Filesystem Architecture

Nexus

Documentation • Quickstart • PyPI • Examples

AI-native filesystem for building intelligent agents

Nexus is an AI-native filesystem that unifies files, memory, and permissions into one programmable layer. Build agents that remember, collaborate securely, and scale effortlessly from prototype to production—without changing code.

Made to build AI agents, fast.

TL;DR

Nexus is a programmable filesystem for AI agents. It combines storage, memory, and permissions into one layer so agents can securely remember, collaborate, and evolve over time.

Jump to: Hello World • Server Setup • Permissions • Core Concepts • Documentation

Quick Start

Installation

pip install nexus-ai-fs

Hello World (10 Seconds)

import nexus

# Zero-config start - data stored in ./nexus-data by default
nx = nexus.connect()

# Write and read a file
nx.write("/workspace/hello.txt", b"Hello, Nexus!")
content = nx.read("/workspace/hello.txt")
print(content.decode())  # → Hello, Nexus!

nx.close()

✅ No configuration needed! Data is stored locally in ./nexus-data/. Mount S3, GCS, or other backends later via configuration.

Embedded Mode (Full Features)

import nexus

# Zero-config local filesystem with AI features
nx = nexus.connect(config={"data_dir": "./nexus-data"})

# File operations
nx.write("/workspace/doc.txt", b"Hello Nexus")
content = nx.read("/workspace/doc.txt")

# Agent memory
nx.memory.store("Python best practices learned from code review")
memories = nx.memory.query(user_id="alice", scope="project")
# Returns: [Memory(content="Python best practices...", timestamp=...)]

# Workflow automation - events trigger automatically!
from nexus.workflows import WorkflowAPI, WorkflowLoader
workflows = WorkflowAPI()
workflows.load("invoice-processor.yaml", enabled=True)
nx.write("/uploads/invoice.pdf", pdf_data)  # Workflow fires automatically!

# Semantic search
results = nx.semantic_search("/docs/**/*.md", query="authentication setup")
# Returns: [
#   SearchResult(path="/docs/auth.md", score=0.89, snippet="...OAuth setup..."),
#   SearchResult(path="/docs/security.md", score=0.82, snippet="...API keys...")
# ]

# LLM-powered document reading (async)
import asyncio
answer = asyncio.run(nx.llm_read(
    "/docs/**/*.md",
    "How does authentication work?",
    model="claude-sonnet-4"
))
# Returns: "The system uses JWT tokens with refresh token rotation..."

nx.close()

Server Mode (Production)

Start the server:

# Configure PostgreSQL backend
export NEXUS_DATABASE_URL="postgresql://user:pass@localhost/nexus"

# Initialize with authentication
./scripts/init-nexus-with-auth.sh

# Server runs at http://localhost:8080

Use the CLI remotely:

# Load admin credentials
source .nexus-admin-env

# Create workspace
nexus mkdir /workspace/project1 --remote-url $NEXUS_URL

# Create users and grant permissions
python3 scripts/create-api-key.py alice "Alice's API key" --days 90
# Output: sk-alice_abc123_...

nexus rebac create user alice direct_owner file /workspace/project1 \
  --tenant-id default --remote-url $NEXUS_URL

# Alice can now access with her API key
export NEXUS_API_KEY='sk-alice_abc123_...'
nexus write /workspace/project1/data.txt "Alice's data" --remote-url $NEXUS_URL

Use the SDK remotely:

import nexus

# Same API, remote execution
nx = nexus.connect(config={
    "remote_url": "http://localhost:8080",
    "api_key": "sk-alice_abc123_..."
})

nx.write("/workspace/project1/data.txt", b"Remote write")
content = nx.read("/workspace/project1/data.txt")

Permission System Example

# Grant different permission levels
nexus rebac create user bob direct_editor file /workspace/project1 \
  --tenant-id default --remote-url $NEXUS_URL
# Bob can read/write

nexus rebac create user charlie direct_viewer file /workspace/project1 \
  --tenant-id default --remote-url $NEXUS_URL
# Charlie can only read

# Check permissions
nexus rebac check user charlie write file /workspace/project1 --remote-url $NEXUS_URL
# Output: ✗ DENIED

nexus rebac check user charlie read file /workspace/project1 --remote-url $NEXUS_URL
# Output: ✓ GRANTED

# Explain permission paths
nexus rebac explain user bob write file /workspace/project1 --remote-url $NEXUS_URL
# Shows: direct_editor → editor → write

Core Concepts

Workspace - A versioned directory for agent state. Create snapshots, rollback changes, and reproduce any historical state for debugging.

Memory - Persistent agent knowledge stored as files. Query memories semantically, consolidate learnings automatically, and share across agents within tenants.

Semantic Search - Vector-based search across files and memories using natural language queries. Powered by pgvector or sqlite-vec.

ReBAC - Relationship-Based Access Control inspired by Google Zanzibar. Fine-grained permissions with inheritance, multi-tenancy, and delegation.

Why Nexus?

Nexus combines files, memory, and access control into a single programmable layer. Build agents that persist knowledge, collaborate securely, and scale from local experiments to distributed systems.

Key Capabilities:

• AI Memory with Learning Loops - ACE system automatically consolidates agent experiences into reusable knowledge
• Database as Files - Access PostgreSQL, Redis, MongoDB through unified file interface with backend-aware permissions
• LLM-Powered Reading - Query documents with natural language, get answers with citations and cost tracking
• Unified Fabric - Files, databases, vectors, and permissions in one programmable layer
• Agent-Native Design - Built for LLM agents and automation frameworks with event-driven orchestration

For AI Agent Developers:

• Self-Evolving Memory: Agents store and retrieve context across sessions with automatic consolidation
• Time-Travel Debugging: Reproduce any agent state with workspace snapshots and version history
• Semantic Search: Find relevant files and memories using natural language queries

For Enterprise Teams:

• Fine-Grained Permissions: ReBAC with backend-aware object types (file, table, row-level access) and multi-tenancy
• Multi-Backend Abstraction: Unified file API for storage (S3, GCS, local) and data sources (PostgreSQL, Redis, MongoDB)
• Content Deduplication: Save 30-50% storage costs with content-addressable architecture

For Platform Engineers:

• Embedded or Remote: Start local (pip install), scale to distributed without code changes
• Complete Audit Trail: Track every operation with built-in versioning and operation logs
• Production-Ready: PostgreSQL backend, API key authentication, and comprehensive observability

Features at a Glance

Category	Highlights
Storage	Multi-backend (S3, GCS, local), versioning, 30-50% deduplication savings
Access Control	ReBAC (Zanzibar-style), multi-tenancy, permission inheritance
AI Intelligence	LLM document Q&A, memory API, semantic search, workspace snapshots, time-travel debugging
Developer UX	Embedded/remote parity, full-featured CLI + SDK, 100% feature compatibility
Extensibility	Plugin system with lifecycle hooks, custom CLI commands, auto-discovery
AI Tool Integration	MCP server for Claude Desktop and other AI agents (14 tools)

Performance Highlights

• 4x Faster Uploads - Batch write API for checkpoint and log files
• 30-50% Storage Savings - Content-addressable deduplication
• Instant Queries - Semantic search with vector indexes (pgvector/sqlite-vec)
• Zero Downtime - Optimistic concurrency control for multi-agent access

Key Features

Storage & Operations

• Multi-Backend Abstraction: Storage backends (S3, GCS, local) and data backends (PostgreSQL, Redis, MongoDB) through unified file API
• Backend-Aware Permissions: Different object types per backend (file vs. database table vs. row-level access)
• Content Deduplication: 30-50% storage savings via content-addressable architecture
• Versioning: Complete history tracking with rollback and diff capabilities
• Batch Operations: 4x faster bulk uploads for checkpoints and large datasets

Access Control

• ReBAC: Relationship-based permissions with Google Zanzibar-style authorization
• Permission Inheritance: Directory-based access control with automatic propagation
• Multi-Tenancy: Complete isolation between tenants with namespace support
• API Key Authentication: Database-backed keys with expiration and rotation

Agent Intelligence

• ACE Learning Loops: Autonomous Cognitive Entity system with trajectories, reflection, and automatic consolidation
• LLM Document Reading: Ask questions about documents with AI-powered answers, citations, and cost tracking
• Memory API: Store, query, and consolidate agent memories with automatic knowledge extraction
• Semantic Search: Vector-based search across files and memories using natural language
• Workflow Automation: Event-driven workflows trigger automatically on file operations - no manual event firing needed
• Workspace Snapshots: Save and restore entire agent workspaces for debugging and reproducibility
• Time-Travel: Access any file at any historical point with content diffs

Developer Experience

• Embedded Mode: pip install and start coding—no infrastructure required
• Remote Mode: Same API, distributed execution with automatic scaling
• CLI: Full-featured command-line interface with remote support
• SDK Parity: 100% feature parity between embedded and remote modes

Extensibility

• Plugin System: Extend Nexus with custom functionality via Python entry points
• Lifecycle Hooks: React to file operations (before_write, after_read, etc.)
• Custom CLI Commands: Add plugin-specific commands to the nexus CLI
• Auto-Discovery: Install plugins with pip, no manual registration needed
• Official Plugins: Anthropic integration, Firecrawl web scraping, and more

AI Tool Integration

• Model Context Protocol (MCP): Expose Nexus to AI agents via MCP server
  • 14 tools for files, search, memory, and workflows
  • Claude Desktop integration with authenticated remote access
  • Run ./examples/mcp/quick_test.sh for 2-minute demo
  • See MCP docs for setup guide

Use Cases

AI Agents - Memory API for context retention, semantic search for knowledge retrieval

Multi-Tenant SaaS - ReBAC for tenant isolation, workspace snapshots for backup/restore

Document Processing - Batch uploads with deduplication, semantic search across files

ML Workflows - Checkpoint versioning, time-travel debugging for reproducibility

Extensible Pipelines - Plugin system (Anthropic, Firecrawl) for custom integrations

📚 See examples/ for complete AI agent, SaaS, ML, and plugin demos with runnable code.

Architecture

%%{init: {'theme':'base', 'themeVariables': { 'primaryColor':'#e3f2fd','primaryTextColor':'#1a237e','primaryBorderColor':'#5C6BC0','lineColor':'#AB47BC','secondaryColor':'#fce4ec','tertiaryColor':'#fff3e0','fontSize':'14px'}}}%%
graph TB
    subgraph agents[" 🤖 AI Agents "]
        agent1["Agent A<br/>(GPT-4)"]
        agent2["Agent B<br/>(Claude)"]
        agent3["Agent C<br/>(Custom)"]
    end

    subgraph vfs[" 📁 Nexus Virtual File System "]
        api["Unified VFS API<br/>read() write() list() search()"]
        memory["💾 Memory API<br/>Persistent learning & context"]
        rebac["🔒 ReBAC Permissions<br/>Backend-aware object types"]
        version["📦 Versioning<br/>Snapshots & time-travel"]
        router["Smart Router<br/>Path → Backend + Object Type"]
    end

    subgraph backends[" 💾 Storage & Data Backends "]
        subgraph storage[" File Storage "]
            local["Local Filesystem<br/>object: file"]
            gcs["Cloud Storage<br/>object: file"]
        end
        subgraph data[" Data Sources "]
            postgres["PostgreSQL<br/>object: postgres:table/row"]
            redis["Redis<br/>object: redis:instance/key"]
            mongo["MongoDB<br/>object: mongo:collection/doc"]
        end
    end

    agent1 -.->|"write('/workspace/data.json')"| api
    agent2 -.->|"read('/db/public/users')"| api
    agent3 -.->|"memory.store('learned_fact')"| memory

    api --> rebac
    memory --> rebac
    rebac <-->|"Check with object type"| router
    rebac -->|"✓ Allowed"| version
    version --> router

    router -->|"File operations"| local
    router -->|"File operations"| gcs
    router -->|"Queries as files"| postgres
    router -->|"KV as files"| redis
    router -->|"Documents as files"| mongo

    style agents fill:#e3f2fd,stroke:#5C6BC0,stroke-width:2px,color:#1a237e
    style vfs fill:#f3e5f5,stroke:#AB47BC,stroke-width:2px,color:#4a148c
    style backends fill:#fff3e0,stroke:#FF7043,stroke-width:2px,color:#e65100
    style storage fill:#e8f5e9,stroke:#4CAF50,stroke-width:1px
    style data fill:#e1f5fe,stroke:#0288D1,stroke-width:1px
    style api fill:#5C6BC0,stroke:#3949AB,stroke-width:2px,color:#fff
    style memory fill:#AB47BC,stroke:#7B1FA2,stroke-width:2px,color:#fff
    style rebac fill:#EC407A,stroke:#C2185B,stroke-width:2px,color:#fff
    style version fill:#66BB6A,stroke:#388E3C,stroke-width:2px,color:#fff
    style router fill:#42A5F5,stroke:#1976D2,stroke-width:2px,color:#fff

Backend Abstraction:

Nexus presents everything as files to users, while backends provide appropriate object types for permission control:

• File Storage (Local, GCS, S3): Standard file objects
• Databases (PostgreSQL, Redis, MongoDB): Backend-specific objects (tables, keys, documents)
• Unified Interface: All accessed through the same VFS API (read/write/list)
• Fine-Grained Permissions: ReBAC uses backend-appropriate object types (e.g., grant access to a PostgreSQL schema vs. individual rows)

Deployment Modes

Nexus supports two deployment modes with the same codebase:

Mode	Use Case	Setup
Embedded	Development, CLI tools, prototyping	pip install nexus-ai-fs
Server	Teams, production, multi-tenant	See Server Setup Guide

Technology: Python 3.11+, SQLAlchemy, PostgreSQL/SQLite, ReBAC (Zanzibar-style), pgvector/sqlite-vec

See Configuration Guide for storage backends (S3, GCS, local) and advanced setup.

Documentation

Getting Started

• Quickstart Guide - Step-by-step setup with authentication
• API Documentation - Complete API reference
• CLI Reference - Command-line interface guide

Core Features

• File Operations - Read, write, delete, batch operations
• Permissions - ReBAC, multi-tenancy, access control
• Memory Management - Agent memory API
• Semantic Search - Vector search and embeddings
• Versioning - History tracking and time-travel

Advanced Topics

• Configuration - Backends, environment variables, YAML config
• Remote Server Setup - Deploying Nexus server
• Multi-Backend Usage - Mounting multiple storage backends
• Plugin Development - Extend Nexus with custom plugins
• Development Guide - Contributing to Nexus

Examples

• Python Demos - SDK usage examples
• CLI Demos - Shell script examples
• Multi-Tenant Demo - SaaS platform patterns
• Authentication Examples - API key setup and usage

Security & Configuration

Security Features:

• ReBAC: Relationship-based access control (Google Zanzibar-style)
• Multi-Tenancy: Complete tenant isolation
• API Keys: Database-backed authentication with expiration
• Audit Trail: Full operation logging with versioning

Quick Configuration:

# Environment variables
export NEXUS_DATABASE_URL="postgresql://user:pass@localhost/nexus"
export NEXUS_URL="http://localhost:8080"
export NEXUS_API_KEY="sk-alice_abc123_..."

See Configuration Guide for YAML config, multi-backend setup, and advanced options. See Permissions Guide for detailed security documentation.

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Development Setup

git clone https://github.com/nexi-lab/nexus.git
cd nexus
pip install -e ".[dev]"
pytest tests/

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Roadmap: GitHub Projects

⭐ If you find Nexus useful, please star the repo to support development!

Philosophy

Nexus treats data as a first-class citizen in AI systems. Instead of building around files, agents build around knowledge—unified, permissioned, and queryable.

We believe AI infrastructure should be:

• Intelligent by default - Storage that understands semantics, not just bytes
• Composable - Mix and match backends, plugins, and deployment modes
• Production-ready - Security, multi-tenancy, and observability from day one

License

© 2025 Nexi Labs, Inc. Licensed under Apache License 2.0 - See LICENSE for details.

---

Built for the AI-native era. Docs • PyPI • GitHub