A broker, not a brain
Providers appear in the hub only from a reviewed registry: a record has to be approved, carry a decision note and state its jurisdiction before it can be shown, and business models where FBT would prepay or resell access under a master account are rejected by name. Alongside the registry you can add your own agent — bring-your-own-agent means the card you supply is validated and the tasks go where you pointed them.
Communication uses the A2A pattern — agent card, send message, read task, cancel task. If a provider reports a state the hub does not recognise, the task is shown as reconciling, never quietly promoted to completed. Kill switches exist globally and per provider, and a failed or unavailable agent is displayed as exactly that.
Consent you can read before anything is sent
Before a task is created you see a preview of exactly what will be shared, composed by the same code path the browser and the server share, so the preview and the payload cannot drift apart. What a task contains then passes through the broker to the provider you chose — that is what sending it means, and the preview exists so you can judge it before it happens.
Provider connections use OAuth with PKCE where the provider supports it. Credentials are stored sealed with AES-256-GCM, connections can be revoked, and results come back rendered as text with the provider named — never as injected HTML from a third party.
The hard limits
An external agent cannot sign for you, cannot move funds and receives no spending allowance over your wallet. Anything on-chain still goes through the normal app flows, where the quote and the 0.70% platform fee are shown and your own wallet signs — an agent can prepare and explain, it cannot execute in your place.
FBT Swap takes no commission from the hub and does not resell agent access. Paid tasks through the hub are not enabled; if you hold an account or a paid plan with a provider, that relationship is between you and that provider, under their terms and their jurisdiction — which is why the registry states it before you connect.
How to judge an agent before you trust it
Read what the consent preview actually sends, and prefer the smallest task that answers your question. Check who the provider is, where they operate and what their own terms say about the data in a task. Treat confident output the way this site treats all model output: plausible text is not evidence, and anything address-shaped or number-shaped gets verified at the primary source.
The hub can make a good agent convenient to reach. It cannot make an agent correct, and it does not convert an agent’s answer into a trade recommendation.