For facts that change, it searches, reads the pages, answers with numbered sources, then fact-checks its own answer against them.
Drop in text, Word, PDF or images. Each chat has its own folder; point a chat at a project folder and it works there.
Python in a sandbox that can only write in the chat's folder. Packages install on demand.
Looks at your photos and screenshots. With images installed it creates pictures and edits your own photos, locally.
Keeps short notes on who you are and how you like answers, in files you can read and edit. Switch it off any time.
Download other local models that fit your computer from inside Joshua, keep them up to date, or pick a bigger cloud model for a chat. On macOS 26 Joshua can also use Apple Intelligence, the model built into the Mac.
A built-in API in the common OpenAI format lets your scripts, a robot or other apps use Joshua, tools and memory included, from this computer or your home network. How to use it
Goal mode works until a task is done and checked. Swarm mode asks several agents and merges the best answer. Scheduled tasks run on their own at set times, while Joshua is open and the computer is on. More
Give Joshua a goal instead of a question. It plans, uses its tools, checks the result and keeps going, round after round up to the limit you set, until it marks the goal as done or tells you why it can't. “Read the three PDFs in this chat and write a one-page comparison as a Word file.”
Several agents work on the same request at once, each in its own way: one methodical, one sceptical, one practical. Joshua then compares their answers, checks the facts they disagree on and writes one final answer. Slower, better on hard or open questions. “What are the strongest arguments for and against a four-day week?”
Ask Joshua to do something at a set time: once, every day, every week or every few hours. Each run appears as a new ⏰ chat. “Every morning at 8, summarise the news about artificial intelligence.” Tasks run while Joshua is open and the computer is on.
The model reasons step by step before it answers, and you can read its reasoning. Slower, but better at maths, logic and problems with many parts. Switch it on only when a question needs it.
Fine-tuning nudges a few of the model's billions of numbers with your examples, so it picks up a habit: a tone of voice, a format, your jargon. Joshua walks you through it in the Teach Joshua panel, explaining each step as you go. It runs on your Mac in 15 to 60 minutes, and your examples never leave it.
Your own model, not a pile of documents. Fine-tuning is not about giving the model documents to search for answers: for that, Joshua has your files, its memory and the web. Fine-tuning changes the model itself. Joshua trains a LoRA adapter: a few million new numbers, next to the model's billions, learned from your examples and then merged into a new model that is yours, with its own name in the model menu.
Why small models. A small model knows less and reasons less well than the giants. But it can be retrained on a laptop in an hour, copied and run anywhere, so it can become a specialist in your tone, your formats, your field's jargon.
It teaches style, not facts. Train it on 200 answers about your company and it will sound like it knows your company, but it will invent the details. For facts, Joshua has memory, your files and the web.
joshua-<name>.The model has no memory of its own. Every time you write, Joshua sends it the whole conversation, and the model either answers or asks for a tool. Joshua runs the tool, adds the result, and asks again, up to 25 rounds.
Built with Ollama · qwen3-vl · FLUX.2 klein via mflux · mlx-lm · llama.cpp's converter · Gradio · anonymous search through ddgs. All open source.