v0.1.0 - Security-First Local AI Agent

v0.1.0 # Local Agent v0.1.0 # Local Agent v0.1.0 - Security-First Local AI

First public release of Local Agent, featuring the Local Permission Broker (LPB) - a novel security pattern for safe AI tool execution.

What is Local Agent?

A local AI assistant that can safely use tools (file operations, memory, automation) without compromising your system. Every action goes through a security "airlock" that evaluates intent, issues cryptographic tokens, and maintains an audit trail.

The innovation: "sudo for AI agents" - nothing executes without proper authorization.

---

Key Features

Security-First Architecture

• Permission Broker: Intent-based access control (not just path sandboxing)
• • Cryptographic Tokens: Single-use JWT tokens with 60s expiry
• • Audit Trail: Every action logged permanently with tamper-detection
• • Sandbox Isolation: Agent operates in ~/local_agent/sandbox only

Intelligent Memory

• Semantic Recall: Vector-based conceptual search (DuckDB VSS)
• • Cross-Session Context: Remembers facts across conversations
• • Efficient Embeddings: all-MiniLM-L6-v2 (384-dim, fast local inference)

Auto-Learning

• Pattern Recognition: After 8 approvals in 24h, trusted actions auto-allow
• • Granular Trust: Only learns specific patterns (e.g., "list sandbox files")

Professional CLI

• local-agent chat - Interactive conversation mode
• • local-agent serve - Launch web dashboard
• • local-agent diagnose - System health check

---

Quick Start

# Clone and install
git clone https://github.com/anandkrshnn/local-agent.git
cd local-agent
pip install -e .

# Pull a model (Ollama required)
ollama pull phi3:mini

# Launch dashboard
local-agent serve

Dashboard opens at http://localhost:8000

Performance Benchmarks
Tested on M2 MacBook Pro (16GB RAM):

Model	Params	Speed	Tool Accuracy	Hallucination Rate
Phi-3 Mini	3.8B	28 tok/s	94%	6%
Llama 3.2	3B	31 tok/s	92%	8%
Qwen 2.5	7B	19 tok/s	97%	3%

Documentation

• Architecture Deep Dive

• • FAQ
• • Contributing Guidelines
    Known Limitations
    This is a working prototype (v0.1.0):

• Single-user only (no multi-tenancy)

• • Memory not encrypted at rest (roadmap: v0.2.0)
• • Limited tool set (4 core tools)
• • Local models only (no cloud LLM support)
    Use for: Personal productivity, learning, experimentation

Roadmap
v0.2.0 (Target: 6 weeks)

• Encrypted memory vault (AES-256)
• • Web search tool (DuckDuckGo)
• • PDF parsing tool
• • Multi-model switcher UI
    v0.3.0 (Target: 3 months)
• Multi-user support
• • Plugin system
• • Mobile companion app
    Part of Sovereign AI Ecosystem
    Local Agent demonstrates security principles that scale to enterprise via:
• PTV Protocol - Open specification for cryptographically-verifiable AI agent identity
  Learn More:
• PTV Protocol Specification
• • Sovereign AI Ecosystem
    License
    MIT License - Free to use, modify, and distribute.