Local AI systems 2026

Nexus / Hermes

A private AI lab where the models run on my own hardware. I use it to find out what agents can safely do without a person watching, and where that line should sit.

Active

The idea

Anything with real consequences stops and waits for a person.

Nexus / Hermes — public system view

01 Capture Preserve the source material

02 Distill Rewrite it without losing where it came from

03 Assign Route bounded work to agents

04 Inspect Show what the agent did and what changed

05 Approve A person approves anything that matters

Agents need an operating environment

Nexus runs local model serving, retrieval, transcription, and task routing in one place. Scout, Librarian, and Forge are separate agent roles, which keeps their work easier to inspect and constrain.

Learning context without losing the source

The question I keep coming back to is how an assistant can learn your context without quietly becoming the only copy of it. So the raw material stays, the source of each claim stays attached, and corrections are possible after the fact.

Small and observable first

A workflow here is deliberately small: capture a conversation, keep the raw record, draft something modest, show what changed between the two, and wait for approval before any of it is treated as real.

What this work demonstrates

Evidence from the work.

A cold-start defect made explicit.
A path mismatch in the orchestrator's file-write tool was traced to how the harness resolved its workspace and recorded as a known operating boundary.
Model choice is measured.
Models are compared on the same RTX 3090 for memory residency, throughput, and task quality. Measured performance determines what earns a place in the lab.
A gateway between agents and the internet.
External search crosses a private, allowlisted gateway with operator status and a kill switch. The gateway was commissioned and frozen before agents could depend on it.
Roles with real permissions.
Named agent profiles hold different access. Research and drafting can be delegated while consequential writes remain outside those roles.
A human approval boundary.
Consequential writes stop and wait. Human approval is a permanent design position for the system.

Public boundary

This is an active lab, not a finished product. What is public here is the design questions and how the thing is run, not a claim that it is ready for anyone else.