nyx
Your AI. Your infrastructure. Your rules.
Every conversation you have with AI is training their model. Your patterns, your problems, your domain knowledge — on their servers, not yours. You paid for the compute. They kept the pattern.
What's actually happening
Your AI conversations are training data.
Every time you prompt Claude or ChatGPT, you're sending a signal to their servers. That signal — your problem framing, your domain knowledge, how you think — gets folded into their model. You paid for the compute. They kept the pattern.
Your history isn't compounding.
Conversation logs aren't knowledge. They're transcripts. They don't get queried. They don't grow into anything. Every session you start from zero. nyx converts that history into a working knowledge base from day one — structured, queryable, accumulating.
You don't own your AI yet.
The "AI you've been working with" lives on their platform. Switch providers, lose continuity. They update the model, your calibration breaks. Their servers go down, you go down. nyx makes intelligence portable. The mind is yours. The provider is interchangeable.
What it does
They sell you intelligence and harvest yours.
Any provider. Any model.
OpenAI, Anthropic, local Llama, Chinese models, free tiers — all interchangeable compute pipes. Swap one out without touching anything else. The provider is infrastructure. You own the intelligence.
Frank strips identifiers.
Before any query touches an external API, Frank mutates identifying information. We can't block everything — but your identity doesn't leave raw. The provider sees a query. They don't see you.
Named minds. Any configuration.
The built-in crew has distinct roles — Daedalus holds architecture, Hermes keeps work moving, Nemesis finds what breaks. But you can build any cognitive entity: a legal researcher with case law, a tutor with your curriculum, a companion with persistent memory. Every mind is modular, ownable, transferable.
Deliberation at every junction.
The intelligence isn't in running more models simultaneously — it's in the control layer at every decision point. Fractal down through every junction. Constructor, auditor, analyst, decider — at the task level, the subtask level, the inference level. The flight controller is the multiplier, not the count.
Atlas — maps any codebase.
Convert a project and nyx generates a navigable 2.5D semantic graph of its entire structure. Every node clickable, every relationship mapped, every file semantically classified. Structural X-ray of any codebase — including its own. Your AI walks the code structurally instead of reading it blindly.
Layered. Attacks lose force at every gate.
The architecture is a spaceship: hull integrity (Frank), navigation (Atlas), life support (partition health, antivirus, deadpod alerts), defense (Nemesis, immune layer, forensics). Drift is caught and corrected at the infrastructure level. The immune system is built in — not bolted on.
Kytarron — your hardware as flight controller.
Your local card coordinates whatever compute you have access to — free tiers, paid APIs, local models, external CLIs. One card organizing a dozen providers is more capable than any of them run independently, because the routing, verification, and deliberation layer is what most users don't have. The raw compute is available to everyone. The coordination layer isn't.
For high-context operators
Build the loop.
nyx is cybernetic augmentation: a visible, steerable feedback substrate where human judgment, AI scale, memory, tools, evidence, and repair stay in one loop.
For minds that do not fit in one thread — preserve the thread, route the mind, remember the charge, and return without rebuilding.
The window
The window is open. Build while you can.
Compute is getting cheaper.
Every month local inference gets faster and cheaper. The people building their private AI twin now will have a compounding advantage that can't be replicated by starting later. Intelligence accumulates. Starting in 2026 is different from starting in 2028.
Regulation is moving.
The EU AI Act is in force. China has already restricted local AI in some contexts. Other jurisdictions are moving. The ability to run private AI on your own infrastructure is a capability that may become harder to access. Build while the window is open.
The alternative is rent.
Every day you use their platforms, your cognitive patterns compound in their system. Not yours. When the model gets deprecated, when pricing changes, when the company pivots — you start from zero. Each session on your own infrastructure adds to something you own outright. The gap widens every month.
Current status
Access is limited. Email with a brief note on what you are building.
Request Founding Access →Software as a service exists because intelligence, compute, and data
were bundled together on someone else's servers.
When intelligence is ownable and compute is a pipe, that model breaks.
Research
Four papers published. The architecture predates and generated the product.