Agents and engines

An agent

An agent is a name, instructions and an engine, stored in your vault. Its conversations stay in the vault too. Nothing about an agent is sent anywhere except the turn you ask it to take.

Local engine

Small open models run inside the tab with WebGPU (WebLLM): Qwen2.5 0.5B (about 400 MB), Llama 3.2 1B (about 880 MB) and Qwen2.5 1.5B (about 1.1 GB). The weights download once and stay in the browser cache. After that, a local turn costs no token and sends nothing over the network.

Relay engines

A relay agent takes its turns at the relay: fast on Claude Haiku 5.5 for one token, deep on Claude Sonnet 5.5 for four. Give it the fetch tool and it can read up to two public pages per turn through the relay. See Spending for what the relay checks and keeps.

Memory

Memory is the transcript the agent keeps in your vault, sent back with each new turn. Clear it and it is gone; there is no copy on a server to clear.