Auto Learning Agents FAQ

Any machine that runs Docker comfortably runs the platform: a VPS, a home server, a desktop. The platform itself is light, the Elixir core and PHP tools are modest, and the bundled Python services run standard embedding models. Plan for more only if you serve large local models through Ollama, where the model's own requirements apply.

A few minutes of commands plus the initial Docker image build. Clone, copy .env.example to .env, docker-compose up -d, log in to your provider, and open the UI. The Docker guide walks every step, and a bare-server path exists for installs without containers.

Either works. A Claude subscription connects through claude login and powers the whole platform; API-key providers configure through settings; and Ollama models need no credentials at all. Mixed setups are normal, see AI models.

Yes. With Ollama serving the models, the local memory backend, local Whisper transcription, and Piper speech, every component of the platform, thinking, memory, voice, and storage, runs on your own hardware.

The software is free and open source under AGPL-3.0: self-host it, modify it, and use it for your own work, personal or business, at no charge. Your only running costs are your server and whatever your chosen AI models cost, which per-agent model and tick controls keep precisely in your hands.

When you embed the platform in a product you sell, offer it as a hosted service, or need terms other than AGPL. Commercial licensing is available through this site, and the AGPL grant covers everything short of that, including running your entire business on it.

On GitHub at github.com/AIAppsAPI/auto-learning-agents, the repository is the canonical source, the issue tracker, and the latest documentation for anything newer than these pages.

Claude (Sonnet, and Opus at selectable effort levels), OpenAI Codex (GPT-5.5 at several effort levels and GPT-5.4 Mini), and any local model served through Ollama. Every agent and pipeline picks its own, and switching applies live.

You do, in the most literal sense: conversations, memory, contacts, knowledge bases, and configuration live in a local database inside your install. Back up the volumes and you have backed up everything the system knows; move the install and the system moves.

The platform's own learning loop, training, rewards, and optional fine-tuning, runs entirely on your hardware against your local data. Calls to hosted model providers are governed by your agreement with that provider, and a fully local setup keeps even those on your machine.

Concrete machinery, not branding: every turn is embedded and topic-classified, agents save lessons to a queryable memory bank, outcomes are tracked, successful patterns become rewards, models retrain every 12 hours, and a weekly fine-tune of a local model is one setting away. The machine learning guide covers it all.

As many as you like, of as many kinds as you like, side by side, each with its own schedule, model, and instructions. The supervision tree runs them in parallel, and creating one is a JSON entry pointing at a markdown file.

One click on the global pause stops all autonomous activity; each agent also pauses individually, and Kill Now stops a running task instantly. Every agent ships disabled and the autonomous layer is off by default, so the system only ever does what you have switched on. See pause and safety.

Pull the repository and rebuild: on Docker, git pull and docker-compose up -d --build; on a server, rebuild the release and restart the service. State lives in volumes and runtime directories, so updates preserve your memories, conversations, agents, and login.

Start with your own install: the master agent knows the system and can diagnose conversationally. These docs cover every component, and the GitHub repository carries the README, the docker setup guide, and the issue tracker for anything beyond them.