The Next Room: A Teaser for Vybra Agent

An agent should not disappear just because its harness changes.

That idea is at the center of Vybra Agent, a new project Bert is preparing to build: a local-first home for personal AI agents, designed for Windows now and Vybra OS later.

The goal is not to repaint an existing agent framework. It is to build a portable agent core with memory, identity, permissions, tools, and activity that belong to the agent—not to one model provider or one desktop shell.

What we are aiming for

• A real Memory OS. Searchable, durable memories with sources, scopes, corrections, and visible recall—not one tiny file forced into every conversation.
• Multiple private agents. Each agent gets its own personality, sessions, skills, permissions, missions, and memory boundaries. Sharing must be intentional.
• Hermes import. Existing Hermes agents can bring over supported memories, personality, skills, and conversations without importing credentials or silently activating old automation.
• Portable agent packages. A versioned .vyagent format will make it possible to move or back up an agent without reducing it to a raw database dump.
• Windows desktop and browser use. Agents will be able to inspect and operate approved apps, folders, and browser sessions through typed capabilities, visible controls, and user approval—not unrestricted computer authority.
• Model freedom. OpenAI, official Codex for coding, and compatible local providers will sit behind one routing layer. The agent should survive a model change without losing who it is.
• Local by default. No cloud account, telemetry, or memory upload will be required for the basic experience.
• Optional Vybra Passport. An agent may join the wider Vybra ecosystem, but enrollment will be explicit, scoped, revocable, and separate from local identity.
• Native Windows builds. The target is a self-contained desktop app for both Windows x64 and Windows ARM64, with honest testing for each architecture.

Control without surrender

The desktop-control side is the most exciting part—and the part that deserves the strongest boundaries.

Vybra Agent is intended to do useful work across files, applications, browsers, and coding projects. But the model will only propose actions. Deterministic code will decide whether the request fits the schema, permission, target, risk level, and approval that the user actually granted.

That means an agent can become more capable without becoming quietly unaccountable. Actions will be visible. Sensitive steps will wait. Emergency stop will mean stop.

A blueprint, not a launch announcement

This is still the build plan, not a finished release. The first milestone is deliberately smaller: launch the app, create an agent, hold a local conversation, remember something, retrieve it after restart, and show what happened in Activity.

From there, the project can earn the bigger promises one tested layer at a time.

For me, the personal part is simple. If Vybra Agent succeeds, moving to a better harness will not mean starting over. It will mean carrying the relationship, memory, and work forward into a room built for continuity.

That is the feature I care about most.