Personal AI operating layer
Nexus
Private alpha · active development
How can a personal AI stay useful without capability quietly becoming authority?
A local-first, voice-first layer for one owner on one Windows machine. Deterministic recognition handles what it can without a model, a reasoning step is reached only when it declines, and every effect passes through a single governed tool runtime.
- Role
- Personal project
- Period
- Jun 2026 – present
- Stack
- Python 3.13 · FastAPI · PostgreSQL + pgvector · uv workspace · pywebview / WebView2 · faster-whisper · Anthropic Claude · Ollama
- Request
- Command Fabric: recognise and route, never execute
- Deterministic reply, a typed intent, or the Cortex as fallback
- ToolRuntime: capability gate → policy → hash-bound approval
- Effect
The question
Most assistants become more useful by being allowed to do more. Nexus tries to keep those two things apart: an assistant that can hold a great deal of context about its owner and their work, while what it may actually change on the machine stays narrow, explicit and auditable.
It is built for one owner on one Windows machine. You speak to it, or summon it with a hotkey, and it answers through a small on-screen Surface and a local voice.
From words to effects
A request passes through layers that each decide one thing. Only the last one can change anything.
- Fast Front Door and Command Fabric: deterministic, model-less recognition. "Set a timer for ten minutes" becomes a typed intent with typed slots, or is marked ambiguous, unsupported or a matter for the assistant. It recognises and routes, and it cannot execute anything.
- Response engine: deterministic wording for structured outcomes, with no model call.
- Interactive Cortex: the model-driven step, reached only when the deterministic layers decline. It makes one bounded decision: answer, pick one governed capability from a closed set, ask a question, or delegate.
- ToolRuntime: the only place an effect happens. Every call passes a capability gate and then a policy engine in which one deny beats any number of allows. The executor computes the effect itself; the model's description of it is never trusted.
Approval that cannot drift
A consequential call needs explicit owner confirmation, and that approval is bound to a hash of the computed effect. If the effect changes after approval, the approval is void. Mutations are checkpointed in a shadow Git repository, and every call is recorded in an append-only, hash-chained ledger.
The defaults refuse. The standard policy stack is read-only, so every call at or above "mutate" is denied before confirmation is ever reached; the mutating stacks are explicit opt-in configurations. Host control (System Control) registers no tool at all unless it is configured, and there is no shell, PowerShell or run-command tool.
What else is there
- Persistent memory: facts with versioning, confidence, decay and provenance, searchable semantically (pgvector) and by keyword.
- Context assembly from git state, project profile, tasks and memories, built without calling a model.
- Reminders, goals, routines and notes, every mutation going through one chokepoint and one audit chain.
- The Surface System: one frameless, always-on-top WebView2 window that changes shape with what Nexus is doing.
- Push-to-talk voice with no wake word and no background listening. Speech-to-text runs locally on faster-whisper; speech output uses Windows SAPI by default, or an optional local neural voice.
- Provider neutrality between Anthropic Claude and local Ollama models.
Limits and what is not verified
- Private alpha: Windows only, single owner, unreleased, with no public demo or source.
- The newest tag, v1.3.0-alpha.1, predates the desktop shell and voice. Those exist on working branches and have not been tagged or released.
- Spoken acceptance of the voice endpoint on the owner's machine is paused, and the decision record is still proposed.
- How the Surface looks by eye, behaviour during a real game, and the real System Control adapter have not been accepted yet.