FAQ

FAQ

Is the demo real? Yes — uncut. The README's GIF is a single continuous recording: the goal typed on camera, a local vision model driving Firefox on a shell-disabled desktop, sixteen GUI actions, a real result. The real-time recording is attached to the release precisely so nobody has to take that on faith.

Does anything leave my network? Not unless you configure it to. Point the engine at Ollama on your own hardware and models, memory, embeddings (a local sidecar ships in the stack) and telemetry all stay local. Cloud providers are opt-in per engine.

Why is the pip package open-agent-fleet but the import agentfleet? agentfleet on PyPI belongs to an unrelated 2025 project. The distribution name differs; the import and the fleetctl command keep the natural names. Same pattern as beautifulsoup4 / bs4.

How safe is it to let an agent loose? Read docs/SECURITY.md — it states plainly what the container boundary is and is not worth, how sudo is gated, and what an egress policy does and doesn't stop. For hostile workloads use the The-QEMU-VM-Tier: a real hardware(-emulated) boundary, with egress enforced where the guest can't reach it.

What models work best? Vision-capable ones — the agent looks at screenshots. Locally, Ollama vision tags; in the cloud, any current OpenAI/Anthropic/Gemini model. Chains let you mix: cheap local model first, cloud fallback for hard turns.

Can I teach it a task instead of prompting it? Yes — that's record-by-demonstration. Do the task once on the agent's desktop with the recorder on; keystrokes, clicks and the accessible element behind each one compile into an editable SKILL.md the agent follows and refines.

What happens when an agent gets stuck? It escalates like an employee: the task parks, your phone gets an alert with the screen frozen at that moment, and you answer from the lock screen — or open the stream and fix it by hand. Nothing silently spins.

Does it scale beyond one machine? One orchestrator drives one Docker host today. Fleets of bots on that host, yes — swarms, pipelines, peer messaging are all built. Multi-host is roadmap; see docs/ROADMAP.md.

What's deliberately not built? The README's Design limits, stated plainly section is the authoritative list, kept current on purpose. Highlights: no GPU passthrough on the VM tier yet, and egress hostname rules resolve once at policy time.

License? MIT. Everything — orchestrator, console, sandbox, SDK, phone app.