OpenParallax

Open-source AI agent security — from a complete personal agent to composable modules you can drop into any stack.

The LLM reasons in a kernel-sandboxed process. A separate process evaluates every proposed action through a 4-tier security pipeline before executing anything. Use the whole agent, or import just the pieces — Shield, Memory, Audit, Sandbox — in Go, Python, or Node.js.

---

A Complete AI Agent

A single static binary that runs on your machine. CLI, glassmorphism web UI, WhatsApp, Telegram, Discord, Signal, iMessage. 69 tool actions across files, git, shell, browser, email, calendar, and more. Semantic memory that persists across sessions. Custom skills loaded on demand. Sub-agents for parallel task delegation. Linux, macOS, Windows — zero runtime dependencies.

curl -sSL https://get.openparallax.dev | sh
openparallax init
openparallax start

A Composable Security Stack

Every security module is a standalone Go package with no dependency on the rest of the system. Cross-language support via Python and Node.js wrappers.

Module	What It Does	Standalone Use Case
Shield	4-tier security pipeline — policy, ML classifier, LLM evaluator, human approval	Drop into any agent framework, or run as an MCP security proxy
Memory	Semantic search — FTS5 full-text + vector similarity, pluggable backends	RAG pipelines, knowledge bases, chatbot memory
Audit	Tamper-evident logging — append-only JSONL with SHA-256 hash chains	Compliance trails for any system
Sandbox	Kernel-level process isolation — Landlock, sandbox-exec, Job Objects	Isolate any untrusted child process
Channels	Multi-platform messaging adapters	Connect any agent to WhatsApp, Telegram, Discord, Signal, iMessage

import "github.com/openparallax/openparallax/shield"

verdict := pipeline.Evaluate(ctx, &shield.ActionRequest{
    Type:    "execute_command",
    Payload: map[string]any{"command": userInput},
})

---

Repositories

Repository	Description
openparallax	The agent — engine, Shield pipeline, web UI, all composable modules, eval suite, and documentation
openparallax-classifier	ML classifier sidecar — DeBERTa ONNX binary for Shield Tier 1. Model on HuggingFace
skills	First-party skill packs — developer, writer, researcher, sysadmin, sales outreach, support agent
mcp	First-party MCP servers — single Go binaries, zero dependencies
parallax-paper	Parallax: Why AI Agents That Think Must Never Act — the research paper behind the architecture

---

The Research

OpenParallax is a reference implementation of the Parallax paradigm, presented in Parallax: Why AI Agents That Think Must Never Act (PDF, arXiv forthcoming). The paper argues that prompt-level guardrails are architecturally insufficient for agents with execution capability — the system that reasons and the system that acts must be structurally separate processes, with an independent security validator between them that neither can modify.

---

Documentation · Contributing · Security Policy

Apache License 2.0