Features
Creobot reads your pages and files, retrieves the relevant passages when someone asks, phrases an answer from them or refuses, hands over when it should not guess, and logs what it could not answer. These pages explain how each part works. Where something is not built, they say planned.
These are not six independent products bolted together. They are one path, and the order matters because each step constrains the next.
Source selection decides what can ever be retrieved, which makes it the highest leverage decision on the list and the one most often made in the first ten minutes without thought. Retrieval decides which passages reach the model. The model phrases an answer from those passages or says it cannot, which is grounding doing its job rather than the model being weak. Where it should not guess, escalation moves the conversation to a person with the thread intact. And every refusal is recorded, which is what closes the loop back to sources.
The practical consequence is that upgrading one part rarely fixes a problem caused by another. A better model does not repair a bad source list. More sources do not improve an answer that failed because the right passage was ranked fourth.
Sources
Creobot takes two kinds of source: pages on your site, reached by crawling, and files you upload. Both are converted into passages, each passage is embedded, and the results are stored in an index that can be searched by meaning rather than by keyword.
Answering
Routing decides which model answers a given question. The decision is made on difficulty, cost and availability, and it includes a fallback path for when the first choice is unavailable, rate limited or too slow.
Answering
Handoff moves a conversation from the assistant to a person, carrying the full transcript, the retrieval trail, the trigger that fired and the page the visitor was on.
Deployment
A single script tag loads the assistant on your site. It is platform agnostic because it is just a script, so it works anywhere you can add one.
Learning
Every conversation is logged with the question, the answer, the passages retrieved and whether the assistant refused. Aggregated by topic, that becomes a picture of where your content works and where it does not.
Trust
Three layers: source control decides what it can retrieve from, grounding and refusal decide what it will say, and escalation rules decide what it must hand to a person regardless.
None of it is publicly available. Creobot opens to teams in order, through the waitlist rather than open signup. Where a specific capability is not built, the feature page says planned rather than describing it in the present tense.
They are one path, not six products. Sources are indexed, a question retrieves passages from them, a model phrases an answer or refuses, uncertainty escalates to a person, and every conversation feeds the log that tells you what to fix. Control sits across all of it.
A feature page explains how a capability works. A use case page explains when a particular team needs it. Handoff and training have both, and each links to the other.
Source selection, by a distance. Everything downstream depends on the right passage being retrievable, and no amount of model quality compensates for indexing the wrong pages.
In practice yes, since routing and analytics are automatic and handoff is configuration. But an assistant without escalation rules is an assistant that guesses when it should stop, so that one is not really optional.
Zapier, Slack and webhooks are planned from the Pro plan. There is no native integration with a specific helpdesk or analytics product, and a script embed is not going to be called an integration.
Limits are, and they are set out in the plan comparison table. Behaviour such as grounding and refusal is the same on every plan, because an assistant that invents answers on the cheap tier would be worse than no assistant.
Signal
Ask. Creobot opens to teams in order, and we will tell you plainly what exists today and what is planned.