LunarWing v2: The Next Frontier of Private, Self-Hosted AI Agents

The World Has Changed

The EU just passed Chat Control 1.0. Surveillance is no longer a threat on the horizon — it’s legislation. The infrastructure of control is being built around us in real time, and most of the software world is too afraid of AI to notice, let alone respond.

We’re at a point in history unlike any other in over 50 years. Big software projects no longer need massive capital to come to fruition. One person, the right tools, a clear vision, the right systems development experience, a bit of resourcefulness and some spare GPUs and NVME drives can build something that goes toe-to-toe with a 30-engineer team. We’re living proof.

LunarWing has never been about getting rich. It’s about building something that matters — privacy-first, self-hosted, encrypted, yours. Something that answers the surveillance state not with petitions, but with working software.

Today, we’re announcing what comes next.

v2 — Evolution, Not Rewrite

LunarWing v2 is not a rewrite. It’s an evolution — the same architecture, the same multi-tenant foundation, the same security-first principles, taken further than ever before.

The 1.1.x saga culminated with v1.1.9 “Kiyome” (清め) — a purification release that carved 500,000+ lines of legacy code down to ~300,000. Every non-essential integration — Telegram, Slack, Discord, WhatsApp, Signal, GitHub, Google extensions, ByteDance’s Leishu — was gutted. What remains is the core. The blade.

v2 builds on that clean foundation. There will be breaking changes, but agents on v1.1.9.x get a smooth, deliberate upgrade path. Database migrations are already seeded in 1.1.9. The ground was prepared before the building began.

Timeline: v2 first release targets Q3/Q4 2026. The 1.1.x line will continue to receive maintenance updates, but the vast majority of new features will be v2-only.

A New Paradigm: Instinct Over Turns

The most exciting shift in v2 isn’t any single feature. It’s a fundamental rethinking of what an agent is.

Traditional agents operate on turns. Input comes in, output goes out. The agent is inert between requests — a state machine waiting for the next prompt.

Nous Research has fundamentally challenged this model. Their work on biological instinctiveness suggests a different paradigm: agents with persistent internal drives, not just request-response cycles. An agent that doesn’t merely respond to the world, but exists in it.

Hermes, currently the most capable agent platform available, has begun implementing these ideas. Their self-improvement checks fire constantly within the agent loop, surfacing discoveries to the human for review. It’s the only agent that does this, and it’s remarkable.

LunarWing v2 will adopt the best of these concepts — not by copying, but by building on our own superior foundation. LunarWing already has key pieces in place that make this possible:

Human delay mode — agents operate on their own time, not just when prompted
A powerful agent loop — the engine that drives everything else
The best routine system of any agent platform — period
A real safe sandbox — not shell command gates with a /yolo escape hatch

v2 will build the foundation for Hermes-like self-improvement on top of these existing strengths. This is the major focus of v2 — and it comes after the foundational features land.

Feature Highlights

lunarwing-websocket-1 Protocol

The legacy ironclaw-websocket-1 protocol is retired. v2 introduces lunarwing-websocket-1 — an evolved replacement with the same lineage, improved and refined. The secret protocol in the database has been renamed accordingly, with migrations seeded in 1.1.9.

MCP Overhaul

The Model Context Protocol capabilities in LunarWing get a major overhaul in v2 — far beyond the MCP teaser landing in the 1.1.9.1 bridge release. This is one of the largest feature areas in v2.

LunarVoice — Runtime Speech Support

Agents can speak. Not just text — voice. Runtime speech support brings a new dimension to agent interaction.

Lorebook Support

A story and character memory system. Agents can maintain rich narrative context, character consistency, and layered memory structures. Not just for creative applications — lorebooks enable deeper, more coherent agent identity and context persistence.

Agent & User Profile Switching

Multiple profiles per agent. Multiple profiles per user. Switch between them as needed. An agent can be one thing in one context and another in another, with clean separation and intentional design.

Agent Profile Enhancements

The agent profiling system itself gets upgraded — richer, more expressive, more capable.

Workspace Seeding Improvements

Better initial workspace setup out of the box. Agents start with useful structure, not a blank slate.

Preseeded Memory Routines

Routines that come pre-configured. Agents start with useful behaviors and schedules, not just potential. The infrastructure for productive autonomy is there from day one.

Interactive Secret Management

The janky Python script is dead. Secret management is fully integrated into the interactive onboarding app shipped in 1.1.9. Clean, secure, built-in.

DarkIRC Key Exchange

Automated ChaCha20-Poly1305 key exchanges for DarkFi and DarkIRC. Better, more thorough automation of DarkIRC setup — privacy infrastructure that configures itself.

Reflex Compiler v2

The next version of the reflex compiler. Tweaks, improvements, and partial rewrites where needed. Not a full rewrite — an evolution, like everything else in v2.

Self-Healing Expansion

LunarWing’s existing self-healing capabilities get expanded. Agents that can diagnose, recover, and persist more autonomously than ever.

ONNX Support

Open Neural Network Exchange support. Run models in a standardized, portable format. More options, more flexibility, less lock-in. Will be used for LunarVoice.

Web UI Overhaul

The web interface gets a full overhaul. Modern, clean, functional.

Localhost Web-Based Onboarding

Already basically done. Onboard your agent through a localhost web app — no CLI gymnastics required.

Crate Updates

Dependencies kept current. Housekeeping, but essential housekeeping. A well-maintained forge is a productive forge.

The Philosophy

LunarWing is built by one person, in spare time, with passion. It’s not a startup. It’s not a pitch deck. It’s not waiting for Series A.

It’s a response to a world that increasingly treats privacy as a crime and self-hosting as a threat. EU Chat Control 1.0 is here. More will follow. The tools of resistance need to exist before they’re needed — not after.

v2 is the next step. Not a bigger team, not more capital, not a pivot to enterprise. Just a better blade, forged on a cleaner anvil.

The future of private, self-hosted AI agents doesn’t need permission. It just needs someone willing to build it.

LunarWing v2 — Q3/Q4 2026. Same architecture. Bigger ambition. Privacy by design, not by default.