The most interesting AI software being built right now isn't the model. It's the loop around the model.

The model picks a tool. The harness validates the arguments, holds the lock, asks for approval if the call is destructive, runs the tool, captures the output, compresses it back into context, and hands the next round back to the model. Repeat until the task is done — or until the loop decides it can't be.

That loop is the product. Liminal is that loop, made visible, made owned, made yours.

What we're shipping

Liminal AI is a model-agnostic agent harness. Everything you need to run a real agent on real software work:

  • A ReAct loop built with the dull-but-critical infrastructure that keeps agents alive — retries with wall-time budgets, provider circuit breakers, context compression, drift scoring, resumable streaming writes, an approval gate on destructive tools.
  • 142 tools organized into lazy-loaded families: files, shell, git, code intelligence with semantic rename, web search and fetch, headless browser with CAPTCHA solving, Obsidian vault read/write, and a document engine that renders PPTX/DOCX/PDF behind a quality gate.
  • Memory that compounds. Typed notes with hybrid BM25 + vector recall. A recipe library that records successful multi-tool strategies and replays them on similar future turns. Background auto-dream that consolidates session logs into long-term memory.
  • A persona system that generates a complete assistant — voice, behaviors, and a themed web shell — from a single prompt at first run.
  • Two UIs. A polished Ink terminal app for keyboard-driven work, and an Express + React web app that streams every tool call inline over SSE. Same harness underneath.
  • One-command install via liminal CLI. The wizard walks through provider setup, runs preflight checks, and opens the web UI with persona bootstrap.

Why a harness, not a hosted agent

Three things bother us about hosted agent products.

They're opaque. You can't see the loop. You can't see why a tool fired. You can't see what's in memory. You can't grep the session trace. You just see "thinking..." for thirty seconds and pray.

They lock the model. They picked a model. Maybe a couple. You don't get to swap. You don't get to point at your local LM Studio. You don't get to A/B Claude against DeepSeek on your actual workload.

They take margin on your spend. Every token you generate is marked up. The hosted layer is between you and the provider you'd otherwise be paying directly.

Liminal is the inverse. It runs on your machine. It speaks the OpenAI protocol against any provider. Your tokens go straight to whoever sells them. Every state change is visible in two UIs and a JSONL trace. The codebase is on GitHub under MIT.

The trade

The trade we're making is that you have to run it.

npm install. A provider key. A few minutes of setup. If you want a "click a button on someone else's website" experience, Liminal is not that product, and we don't want it to be.

What you get in exchange is something that does not exist anywhere else: a serious agent harness, fully open, with two real UIs and the infrastructure underneath to keep running through the failure modes that take down most demos in the wild.

How to try it

Follow the full install guide or run:

curl -fsSL https://vireondynamics.com/install/install.sh | bash

Windows:

irm https://vireondynamics.com/install/install.ps1 | iex

The installer handles everything — clone, install, build, wizard, doctor check, web UI launch with persona bootstrap.

Learn more: Liminal overview · Features · Compare vs Cursor & ChatGPT · Pricing (free)

If you have feedback, find a bug, or want to talk: GitHub issues, hello@vireondynamics.com.

We've been quietly building this for months. v0.1 is the foundation. There's a lot more coming — but it's the right time to put it in front of you.

— The Vireon Dynamics team