Our clients started asking for AI on their websites. We went looking for something to give them, and came back empty-handed. So we built it ourselves. It is called copadre, it is live now at copadre.com, and this is what we love about it.

Reliable agents. Silent routing.

Building one assistant that knows everything is a trap. Every time you add to its knowledge, you risk confusing it, and it costs more tokens. The bigger it gets, the less you can trust it — and the more it costs.

copadre is made for creating teams of small, specialist agents. You build your agents, each doing one job properly, and add a concierge agent to silently route visitors to the right one. Agents can also hand back when a conversation drifts outside their remit, or pass someone deliberately forward through a multi-step process.

A small agent has less to think about, so it is more reliable, and it is cheaper to run.

No bluffing. Just learning.

If a concierge keeps getting a conversation passed back, this is a signal that the visitor has asked something none of the agents can handle — and your concierge knows exactly what was asked and when this happened, so it logs it for you. We call this the out-of-knowledge loop and it is how you know what training to add to your agents.

"An assistant that tells you when it needs training will quickly improve. One that guesses responses and never tells you will always be mediocre."

- Matthew Gibson, Director, This Company

When a visitor asks something outside your agents' knowledge, copadre says so plainly, in wording you have written. The moment is flagged on the conversation transcript and highlighted on your copadre dashboard.

Not another app. Use AI to manage AI.

copadre is fully MCP-compatible. This means you can chat to the platform from your favourite AI tool. No need to log in or learn another app. You can connect Claude — or any AI tool that speaks MCP — straight to your copadre account.

Tell Claude to add a new document to a research library assistant and off it goes: locate the assistant, download its agent definition, summarise the new document, save it to a draft agent definition, test the change via preview chat and confirm when it passes. Two minutes later, when you're ready, just tell Claude to publish.

Reporting is just as easy. At the end of the month you ask your AI to read your conversations and report on what people are asking, compare this month to last, pull out any out-of-knowledge events and describe what training has been added. Set it on a schedule and it is in your inbox on the 1st, ready to share with your clients.

Simple pricing with cost controls.

copadre charges according to how many agents you have and takes no markup on AI whatsoever. Your agents run on your own Claude API key and Anthropic bills you directly under your own agreement with them.

We don't sell tokens, but we don't want you to waste them. Conversations that run past a sensible length are paused. Abusive visitors are blocked for a cool-off period you choose. Prompt caching cuts the cost of long conversations. With copadre you can even pick the model you want each agent to use — assign a cheap, fast model to simple agents and reserve the expensive stuff for where it counts.

Ready for the general public.

An assistant on a public website represents your brand and must not misbehave. We assume that people will try to get it to do wrong things, so we built protections against that happening.

Visitor messages are cleaned of anything trying to hijack your agent's instructions or routing. Every link an agent offers is checked and pointed at your own domain — so even a compromised agent cannot send your visitors somewhere unpleasant. Each embed key works with exactly one agent, so a key lifted from your page cannot be used elsewhere. Agents can be locked to your domains so they will only work on your sites. Instructions, conversations and your API key are encrypted at rest, and every agent inherits copadre's own guardrails on top of whatever you have written.