A sovereign macOS agent that lives entirely on your silicon. No telemetry. No subscriptions. No cloud gatekeepers. Your data doesn't leave your machine — ever.
Let's be honest about what most AI tools actually are. They're beautiful UI wrappers around a deal you didn't fully read. You type something private, it travels to a server farm in Virginia, gets processed alongside millions of other people's secrets, and a sanitized answer comes back. The model gets smarter. You get a chat window. They get your data.
You don't own your conversation history. You don't own your embeddings. You don't even own the context window you're renting by the token. Every query you make is a debit from a cognitive checking account that belongs to someone else.
That's not intelligence. That's a subscription to your own subjugation.
Arkhein is the answer to a question most developers were afraid to ask: what if the AI lived on your machine and reported to nobody?
Arkhein is not an AI tool. It is a declaration that your silicon, your storage, and your mind belong to you — and no one else.
Arkhein is a native macOS application — built with NativePHP, the framework that finally let PHP developers colonize the desktop — that runs a complete AI reasoning pipeline entirely on your local hardware. No API keys. No internet dependency. No monthly invoice from a hyperscaler.
At its core, Arkhein is four interlocking systems operating in perfect silence under a minimalist shell. Each one built to answer a single radical question:
Verticalized RAG. Every folder is a sovereign knowledge partition. Project A never contaminates Project B.
Local inference. Mistral for efficiency. Qwen3 for depth. A multi-stage arbiter that thinks before it speaks.
SQLite is the Single Source of Truth. Vektor is the disposable binary accelerator. Self-healing, zero-downtime.
Filesystem commands with surgical precision. You remain the final authority. Always.
Every other RAG system dumps your documents into a single vector pool and hopes for the best. The result? Cross-contamination. Your medical research blending into your client contracts. Your journal entries influencing your product specs. Context rot.
Arkhein refuses this slop. It organizes knowledge into Sovereign Silos — isolated, physical partitions with a strict hierarchy:
Before the agent reads a single fragment, it already knows the document summary and the subfolder depth. This is Parent-Aware RAG — understanding context before retrieving content. The result is retrieval that actually makes sense, instead of a bag of semantically proximate word-soup.
Your silos are yours. Authorize exactly which folders Arkhein can see. The rest of your filesystem is none of its business.
Ollama runs the inference. But what Arkhein does with Ollama is where it gets interesting. It doesn't just prompt a model and relay the output. It runs a Multi-Stage Sovereign Arbiter pipeline — a JSON-based reasoning engine that classifies your query intent before triggering any retrieval.
The Dispatcher decides what kind of query you're asking. Retrieval? Synthesis? Filesystem command? Creative generation? Intent shapes execution. The agent thinks before it acts — a behavior that's surprisingly rare in an industry obsessed with speed over accuracy.
Two compute profiles let you match the hardware to the mission:
| Efficient Profile | Mistral + Nomic Embed — optimized for 8GB–16GB RAM. Fast, capable, honest about its limits. |
| Elite Profile | Qwen3 Suite — high-precision reasoning for 32GB+ RAM. Deep synthesis. Complex multi-document analysis. |
| Polyglot Mirroring | The Archivist responds in the language you write in. Always. No configuration required. |
The models are local. The inference is local. When you shut your MacBook, the intelligence stays there — not in a data center, not in a training pipeline, not in someone else's product roadmap.
You are not a user. You are not a data point. You are the Operator — and Arkhein is your agent, not someone else's.
The memory layer is where most local AI projects break down. Vector databases are temperamental. Embeddings go stale. Re-indexing kills performance. Arkhein solves this with an architecture it calls the Self-Healing SSOT.
SQLite is the Single Source of Truth. It holds your chunks. It holds your document hierarchy. It holds everything that matters. SQLite doesn't fail silently. SQLite doesn't corrupt your data and blame a network timeout. It's a file on your machine — inspectable, portable, yours.
Vektor is the binary accelerator layer — disposable by design. If the vector index goes sideways, Arkhein detects the dimensional mismatch, scraps it, and rebuilds from SQLite without taking the system offline. Your archive stays live during the entire rebuild. Hot-swap reconciliation using shadow partitions. Zero downtime. No panic.
Arkhein can command your filesystem. /create, /move, /organize, /delete — strategic operations with surgical precision. But here's the thing that matters: it never acts without a verified Strategic Plan that you explicitly approve.
This is Human-in-the-Loop as a design principle, not a legal disclaimer. The agent proposes. You decide. Arkhein executes. In an era where agents are racing to be more autonomous, more "agentic," more comfortable taking irreversible actions on your behalf — Arkhein's insistence on your final authority is a radical act.
Your machine. Your files. Your call.
The audacity of Arkhein's stack is half the point. A native macOS AI agent built on PHP and Vue. The web framework mafia said this was impossible. NativePHP said hold their beer.
The entire application is open source under the MIT License. You can read every line of code that makes decisions about your data. You can fork it. You can audit it. You can modify it. Sovereignty isn't a brand promise — it's a build artifact.
Getting it running locally takes four commands:
# Pull the inference engines ollama pull mistral ollama pull nomic-embed-text # Assemble the machine composer install && npm install php artisan native:migrate php artisan native:serve # → Sovereign architecture active. # → No cloud. No telemetry. No compromise.
Beyond the verticalized silos sits the Sovereign Archivist — a centralized module for global RAG and system intelligence. When you need to search across every authorized silo simultaneously, or analyze the architecture itself, the Archivist is your command center.
It's the meta-layer. The intelligence that understands the intelligence. It knows which silos exist, what they contain at a summary level, and how to route your queries for maximum precision. Not just a search box — a librarian who has read everything you've given it and remembers where it all lives.
Arkhein v0.0.4 is available right now for Apple Silicon Macs. It's raw. It's early. It's exactly the kind of software that will look back on this moment as chapter one. The architecture is solid. The vision is fully formed. The edges need filing.
If you want a polished consumer app with a customer support team and a changelog written by a marketing department, this isn't your tool yet. If you want to run a sovereign intelligence system on your machine today, contribute to something that matters, and be part of the community that decides what local-first AI actually means — you're early and that's a feature, not a bug.
The cloud won't give you your data back. But you can stop sending it there in the first place.
Apple Silicon · Zero Cloud · Open Source · MIT License