Agents

An agent is a chat window on your site, answering as your company, from material you gave it.

What it is

One agent is one assistant with one personality, one set of documents and one set of tools. Most companies need one. You would create a second when the audience is genuinely different — a support agent for customers and an internal one for staff read different material and should not share a memory of either.

Each agent gets a script tag. Paste it into your site and the chat launcher appears; everything else — colours, position, greeting, which documents, which tools — is configured in the dashboard and takes effect without touching the page again.

your site

<script
  src="https://cdn.tigerbase.cloud/v2"
  data-api-key="tb_your_agent_key"
  defer
></script>

The key is not a secret and is not treated as one — it sits in your page source where anyone can read it, the same way a maps or payments publishable key does. What makes it yours is the domain list: add the sites this agent is allowed on under Settings → Allowed Domains, and the same key pasted into anyone else's page is refused. Everything the key can do is bounded by your plan's limits, and it can only talk to this one agent — it cannot read your conversations or change anything.

What happens when someone asks something

Retrieval first, then the model, then tools if the model asks for them.

The important part is the first step. The agent does not answer from what the model happens to know — it searches your material, and the passages it finds are what the model is told to answer from. That is why an agent with no documents is nearly useless, and why an agent with the wrong documents is confidently wrong.

What it decides on its own

Which passages to use	Chosen by similarity to the question, not by you.
Whether to call a tool	The model decides, from the tool's description. A vague description is the usual reason a tool is never used.
Whether to ask for a person	Only when you have enabled it. Off by default.
What it will not do	It cannot create agents, change its own settings, read another tenant's data, or run anything you have not given it.

An agent is not a person, and not a database

It produces the most plausible answer given what it retrieved. When your documents are silent on a question, a good agent says so — but it is a model, not a lookup, and the closer two documents are in wording the more likely it is to blend them. For anything where being wrong is expensive, give it a tool that reads the real system rather than a document describing it.

Which model answers

Agents run on Google Gemini. You can use ours, or supply your own API keys — several, if you like: they are tried in turn, so one key being refused or out of quota does not take the agent down.

Your keys are used alone

When you supply keys, we never quietly fall back to ours if they all fail. You are told the key stopped working — by a notification, and in the agent's logs — rather than discovering it on an invoice.

Knowledge — what happens to a document you upload