Luminarys Platform

Luminarys is a sandboxed AI skill platform with multi-node clustering, MCP/ACP support, and a built-in autonomous agent.

Skills run in isolated sandboxes with fine-grained permissions. Write them in any language that compiles to WebAssembly, deploy on any architecture — from cloud servers to IoT devices.

What problems does it solve

Secure AI execution in regulated environments. AI agents need access to file systems, networks, and external APIs — but in enterprise and industrial settings, unrestricted access is unacceptable. Luminarys enforces fine-grained permissions at the ABI level: every file read, HTTP request, TCP connection, and shell command is checked against a declarative policy before execution.

Run anywhere — cloud to IoT. Skills compile once and run on any platform without recompilation: Linux, macOS, Windows — on x86, ARM, RISC-V, MIPS. From cloud servers to Raspberry Pi, industrial controllers, and embedded gateways.

Vendor-neutral AI integration. The platform exposes skills via MCP (Model Context Protocol), making them accessible to Claude Desktop, Cursor, Qwen CLI, and any compatible client. Skills are not tied to a specific LLM provider.

Multi-tenant isolation. Each skill runs in its own sandbox with a separate memory space and permission scope. A compromised or buggy skill cannot access data belonging to other skills or the host system.

Key capabilities

Sandboxed skill execution

• Skills run in isolated sandboxes with their own memory space
• No access to host filesystem, network, or processes except through controlled ABI functions
• Permission policies enforced at every ABI call
• Declarative deployment manifests define what each skill can do

MCP & ACP protocols

• Full MCP support: Streamable HTTP, Legacy SSE, stdio
• Skills appear as typed MCP tools with parameters and structured output
• Compatible with Claude Desktop, Cursor, Qwen CLI, MCP Inspector
• ACP (Agent Context Protocol) support is planned

Multi-node clustering

• Master-slave architecture connected via NATS
• Clients see one MCP server — skills execute on the node that has them
• Cross-node file transfer built in
• Add capacity by adding nodes — no reconfiguration needed

Signed skill packages

• Every .skill package is cryptographically signed
• The host verifies the signature before loading
• No unsigned code runs on the platform

Agent mode (in development)

• Autonomous agent that orchestrates skills to complete complex tasks
• Inter-skill invocation, scheduled tasks, event-driven workflows
• Persistent state management
• LLM-in-the-loop for reasoning and decision-making

Target environments

Enterprise & regulated industries

• Financial services — AI agents with auditable data access
• Healthcare — compliant skill execution with strict data isolation
• Government — air-gapped deployments with sandboxed skills
• Manufacturing — process automation on factory floor controllers

Edge & IoT

• Industrial gateways (ARM, RISC-V, MIPS) — lightweight runtime for constrained devices
• Autonomous systems — local AI processing without cloud dependency
• Smart infrastructure — building management, energy optimization

Cloud & hybrid

• Multi-region clusters with node-level skill distribution
• Hybrid cloud/edge topologies for latency-sensitive workloads
• Configurable resource limits per skill

Skill languages

Skills can be written in any language that compiles to WebAssembly. SDKs are currently available for three languages:

Language	Best for	Binary size
AssemblyScript	Edge/IoT, high-density nodes, constrained environments	~20–40 KB
Rust	Performance-critical workloads, system-level skills	~250 KB – 5 MB
Go	Rapid development, complex business logic	~1–15 MB

All three SDKs share the same ABI — skills from different languages run side by side on the same host.

Security model

Every skill operation passes through the permission engine:

Permission	What it controls	Example
fs	File system access	dirs: ["/data/project:rw", "/config:ro"]
http	Outbound HTTP requests	allowlist: ["https://api.example.com/**"]
tcp	Raw TCP connections	allowlist: ["db-host:5432"]
shell	Command execution	allowlist: ["git **", "go build **"]
file_transfer	Cross-node file copy	allowed_nodes: ["master"]

Additionally:

• DNS-aware network filtering
• Glob patterns supported for flexible path matching

Next steps

• Architecture — how the platform works internally
• Installation — download and run
• Configuration — YAML config reference
• Quick Start — create your first skill