Get started
npx lloyal-ai newOne command creates a working app: a desktop window, a browser app and a terminal, all running the same program over a model that is downloaded and verified on its first run. The program — which agents exist, what they read, when they are done — is ordinary TypeScript in src/harness/, and it is yours to change.
- Quickstart — three commands to a running app.
- System requirements — what your machine needs. New to the terminal? It has a three-step Node.js install.
- Build your first harness — find the program inside the project and change how it thinks.
What you can build
| An app people download — the model built in, working offline, no account. | Ship a desktop app |
| AI inside your product for many users — one model on your own machine or GPU box, a session per user, no per-token bill. | Serve to many users |
| Research you can hand someone — agents that read in parallel from one shared context and settle on a cited answer. | The research template · Agents and orchestration |
| Classification and extraction — a label, a number, a choice from a list, from the model you already have, with nothing to parse. | Structured output |
| Reasoning with specialists beside it — a judge that ranks what agents read, vision, embeddings, named in one line each. | Services · Retrieval |
| Tools that work inside live inference — tools that know what the calling agent has read, and can start agents that inherit it. | Tools |
| Answers from your documents — PDFs and images attached by a reader, searched and shown to the model only where it needs them. | Attachments and documents |
Find your way
| You want to | Start here |
|---|---|
| Run several agents over shared context | Agents and orchestration |
| Give agents an action, or your own data | Tools |
| Refuse a call, retry a failure, require evidence before an answer | Tool hooks and guards |
| Limit turns, time and context — and wrap up early | Agent policy |
| Let an agent act only with a person's approval | Human approval |
| Change what the model is told | Prompts |
| Change or bring your own model | Models |
| Add a setting a user can change live | Settings |
| Build the screen around it | The interface |
| Test it without a model | Testing |
| See what the agents did | Debug with traces |
| Fix an error | Troubleshooting |
| Package a capability for any harness | Abilities |
| Look up a command or a setting | CLI · harness.yml |
How it works
Lloyal programs are built on Effection, Frontside's structured concurrency library: whatever a piece of work starts is finished or cleaned up when that work ends. If you write async/await, Structured concurrency is the translation, on one page.
Every agent is a branch of the model's live state — it forks from what the model has already read, rather than re-sending it — and the pool advances all of them together over one model. Why that changes what an application can do is Continuous Context; how it is programmed is Thinking in Lloyal.
For your coding agent
Every page here is also Markdown — add .md to its URL — and /llms.txt lists them all. Every scaffolded project carries an AGENTS.md: the rules and the map a coding agent needs before it edits the code.
Lloyal is the platform for Vertical Inference.