Skip to content
HUIAHUB

Models & inference

Lloyal

A TypeScript framework for AI apps that run an open-weight model inside the app, offline and without an API key.

Verified Open source
Visit official site (opens in a new tab)

What Lloyal does

Lloyal is a TypeScript framework from Lloyal Labs for building applications that carry their own language model. Running npx lloyal-ai new scaffolds a project with a desktop window, a browser app and a terminal interface, all running the same program. The model is named in a harness.yml file, downloaded on first run, checked against a pinned digest and loaded into the same process as your code, so the app makes no calls to a hosted model API.

The application logic is ordinary TypeScript in src/harness/, and you change it like any other code. Each agent is a branch of the model's live state, so several agents can work from one shared context instead of each loading its own. The docs cover tools, tool guards, human approval, typed decisions for classification, retrieval over attached PDFs and images, and debugging with traces. Lloyal also hosts a signed channel of first- and third-party add-ons it calls Abilities.

The default model is Qwen3.5 4B (2.6 GB of weights). The catalog also has two Qwen3.8 27B builds, a reranker and embedding models, and you can bring your own GGUF file. Requirements are Node.js 24 or newer and at least 10 GB of memory; a GPU is optional, with Metal on Apple silicon and CUDA or Vulkan on Linux and Windows. The ship command packages a macOS disk image, and a served mode loads the model once and gives each browser user a separate session.

Lloyal suits TypeScript developers who want to ship local AI apps or host a private multi-user service on their own hardware. The code is MIT-licensed, and there is no pricing page.

  • Ship a desktop app with the model built in that works offline and needs no user account.
  • Serve one local model to several users, each with a separate session, on a private machine or GPU box.
  • Run parallel research agents over a shared context that return a cited answer.
  • Pull a label, number or choice out of text with the local model, using typed decisions.
  • Answer questions over PDFs and images that a user attaches.

Who should use it

  • TypeScript developers building local or offline AI applications.
  • Teams that want to host one model for several users without per-token billing.

Who should look elsewhere

  • Machines with less than 10 GB of memory, which the installer refuses for the default model.
  • Developers who need signed desktop builds for Windows or Linux today, since `ship` produces a macOS disk image.
  • Teams that work in Python rather than TypeScript.

Strengths

  • Models download once, are verified against a pinned digest and run without an API key, Docker or a vector database.
  • Agents fork from shared model state, so several can work over one context.
  • The same program runs as a desktop app, a browser app and a terminal tool.
  • MIT licence, with the source on GitHub.

Limitations

  • Needs Node.js 24 or newer and 10 GB of memory for the default 4B model; the 27B catalog models need 24 GB.
  • Building a distributable desktop app requires macOS, and the output is a macOS disk image.
  • No pricing page exists, and the cost of third-party Abilities, if any, is not stated.

Sources