Core concepts

A short tour of the words copadre uses, and how the pieces fit together. None of it is essential to get a first agent live — but it makes everything else read more easily.

Organisation and team

A copadre account is a shared organisation — a team workspace, not a single login. Everyone you invite gets their own passwordless sign-in, and each person has a role that decides what they can do.

Individual logins are what make roles meaningful: a clear trail of who built what and who saw what, and clean security if someone leaves.

Agents: concierge and specialist

“Agent” is the umbrella word for one AI chat persona — its instructions, model and settings. Agents come in two kinds:

• A specialist carries out a single role — answering questions, qualifying leads, walking someone through a process.
• A concierge is a front-of-house router that hands each visitor to the right specialist. A larger setup might have several concierges; a simple one needs none at all.

Which kind an agent is gets fixed when you create it. There’s no “default” agent for an organisation — every embed points at a specific agent. When you do run several, see multi-agent routing for how the #slug# hand-off works.

Draft and published

Every agent has two versions: the draft you’re editing and the published one your visitors actually talk to. You change a draft, test it in preview mode, and only when you choose Save and Publish does the live agent move. Your widget never shifts under your customers mid-edit.

Embed keys

An embed key is the token in your widget’s <script> tag. Each key is bound to one agent — copadre creates it for you when you create the agent, and the snippet carries both the key and that agent’s ID. Because the key sits in public HTML, you pair it with domain locking (below) so only your sites can use it. A leaked key exposes one agent, never your whole roster.

Your Claude API key

copadre runs every agent on your own Claude API key. Inference is billed to you by Anthropic directly; copadre never sees your bill. Your key is encrypted at rest. This is the “your AI, your bill” idea, and it’s why the API key step comes first.

Controlling cost and abuse

Because it’s your spend, copadre gives you direct levers over it:

• Per-agent model choice — a fast, cheap model for simple agents, a stronger one only where it earns its keep.
• Prompt caching — an opt-in per-agent toggle that reuses an agent’s instructions between turns, so longer conversations cost less. Switch it on once an agent is settled. Not recommended for concierges: a concierge usually handles a single turn before handing the visitor on, so there’s no follow-up turn for the cache to save on.
• Domain locking — restrict each embed to your allowed domains; requests from anywhere else are refused before any AI call.
• Turn limits & abuse blocking — a configurable cap and automatic blocking stop bots and time-wasters, killing the connection before a wasted call is made.

Conversations and the out-of-knowledge loop

Every chat lands in your conversation dashboard. A conversation counts as finished after a quiet period you set (the cooldown); finished threads can be emailed to you in full. The piece that makes copadre a platform rather than a widget: when an agent can’t answer, copadre surfaces that out-of-knowledge question so you can teach it — then tracks whether your fix worked.

What copadre adds to every agent

You are not writing your agent’s instructions on a blank slate. copadre appends its own instructions to every agent, on every request, so a set of things are handled for you whatever you write:

• How long a reply is, and how many questions it asks. Your agents keep answers to a sensible length and don’t interrogate visitors.
• Formatting. Replies use the small set of formatting the chat widget actually renders, so nothing arrives as raw symbols.
• Links. Agents can only link to pages you have told them about, and links are always kept on your own site.
• What an agent may commit to. Your agents give and gather information. They don’t agree prices, make promises or enter into anything on your behalf.
• Abusive messages. Handled before they reach your agent’s instructions at all.

So you don’t need to write any of that yourself — and it’s better if you don’t, because your own version can only contradict what the platform is already doing. Spend the effort on the part only you can write: what your agent knows about your organisation, and how it should sound.

You can always be stricter than the platform. Telling an agent to ask at most one question, or to keep to a single short paragraph, works fine. It’s loosening these that won’t stick.

Ready to put it together? Walk the quickstart, or grab the starter agent to adapt.