Core concepts

A handful of ideas explain almost everything in SiteTrellis. Read this page once and the dashboard will feel familiar.

Accounts, workspaces, sites and agents#

Everything you create lives in a simple hierarchy:

Conversations, messages and leads belong to a site and an agent. Each site has a default agent; when a visitor opens the widget, that agent (or the first enabled, published agent on the site) answers.

Roles#

There are three roles. See Team and roles for details.

• Owner: everything, including billing and deleting the account.
• Admin: manage sites, agents, keys, webhooks and team members, and see every workspace.
• Agent: work the inbox, reply to visitors, take over and hand back conversations, and view leads. Agents cannot change configuration and only see the workspaces they have been added to.

The word "agent" therefore means two things: an AI agent you configure, and a person with the Agent role who answers chats. The documentation says "AI agent" or "operator" where it matters.

Draft versus published#

Each agent has a working copy, the draft, and zero or more immutable versions. Editing the agent changes only the draft. Visitors on your website are always served by the published version, so you can experiment freely.

• Playground runs the draft.
• Publish checks the draft for problems, then snapshots it as the next version number, with an optional note.
• Rollback re-publishes an earlier version as a new version, so history is never rewritten.
• Each conversation remembers which version answered it, which makes it easy to see whether a change helped.

Your plan limits how many versions are kept; older ones are pruned. More in Playground, publishing and rollback.

Bring your own key#

SiteTrellis does not run AI models for you or resell tokens. You create an account with an AI provider, generate an API key, and store it in SiteTrellis. Each reply your agent writes is a request made with your key and billed by the provider to you. This keeps AI costs transparent and lets you choose the provider and model that fit your budget. It also means SiteTrellis plans can price on the platform itself: conversations, agents, seats and features. See AI keys overview.

An agent's model chain is an ordered list of steps, each a key plus a model. If step one fails (timeout, rate limit, provider outage, invalid key), the next step is tried automatically, so your visitors are less likely to see an error.

Visitors, conversations and messages#

• A visitor is an anonymous browser session identified by a generated id (vis_...). You can attach details to a visitor with the JavaScript API.
• A conversation (conv_...) is one chat. Its mode is ai while the AI answers, human while an operator has taken over, and closed when it has ended.
• A message has a role: visitor, assistant, operator, or system (small centered notes such as "Sam joined the chat"). Tool calls are recorded but never shown to visitors.

Plans and limits#

Plans limit conversations per month, and the number of sites, agents, workspaces, team seats, fallback-chain steps, custom HTTP tools and kept versions. Some features are gated: human takeover, outbound webhooks, post-chat analysis, and removing the "Powered by" label together with custom CSS. A conversation for billing purposes is one visitor session with at least one visitor message. Playground chats are not counted. When you approach the limit the dashboard tells you; see Billing and plans for what happens at the cap.

What runs where#