AgentPlayground is a self-hosted AI agent platform. You run it on your own computer (or your own
server), and it gives you a system of AI agent teams — coordinated by a single central assistant —
that can do real work: research, write, organize files, manage a knowledge base, and delegate
across specialists, instead of a single chat window that only ever talks back.

It runs 100% free two ways: NVIDIA's free cloud models (one free API key, no credit card) or
fully local models via Ollama on your own machine. It also works with Anthropic Claude and OpenAI
if you have paid keys, and you can mix providers — the coordinator on Claude, a team on a free
model — in the same install.

You don't talk to a dozen different tools. You talk to one assistant, and it routes the work.

• The coordinator is that assistant (also called the Playground Keeper in the code and prompts you'll see). It's the only thing you need to open a conversation with. It decides whether a request is a quick one-off task, something to delegate to a specialist team, or a multi-step plan that needs several teams working together.
• Teams & agents are the specialists. A team is a named group (for example "Research", "Blog Team") made up of one or more agents, each with its own model, its own instructions, and its own job. The coordinator delegates to a team by name; the team runs its own tool loop and reports back.
• Skills are reusable instructions a team can be given — an invoice generator, a CRM contact manager, a meeting summarizer, and more. Skills tell an agent how to do a specific kind of task well, on top of its general abilities.
• The Brain is the shared knowledge base. It's a set of notes (a "vault") that agents read from and write to — project briefs, research, meeting summaries, CRM data — searchable and organized by folders and tags. Anything a team learns or produces can be saved there so the next task (or the next team) has the context.
• Playgrounds are containers that group teams together by context — for example, a "Personal" playground with a trainer and a finance advisor, or a "Business" playground with sales and content teams. Each playground can scope which parts of the Brain it sees.
• Meetings let you put two or more teams (or individual agents) in a live, round-based discussion — they debate a topic out loud, in real time, and the coordinator (as facilitator) closes the meeting with a decisions list you can dispatch straight into new tasks.

A typical session looks like: you open a chat with the coordinator, describe what you need, and it
either handles it directly, delegates to the right team, or — for anything bigger — proposes a plan
across several teams. Along the way, results get saved to the Brain so your agents get smarter about
your specific business or life over time, not just smarter in general.

1. [Install the desktop app](/docs/install) — get AgentPlayground running on your machine.
2. [Prompt Templates](/docs/prompt-templates) — copy-paste starting points for the most common requests (create a team, dispatch a task, schedule a job, and more).
3. [Customizing & Editing the App](/docs/architecture-overview) — if you want to go under the hood: how the app is put together and how to change it.