CoreAI Blog
CoreAI updates and articles about AI
One knowledge base, both MCP client and MCP server
coreAI is both an MCP client and an MCP server in the same knowledge base. As a client the knowledge base can not only fetch fresh data but also create, update and delete in other systems on the user's behalf – agentic AI that completes tasks, not just answers. As a server you expose the knowledge base so other AI agents, IDEs and internal workflows can query it over an open, vendor-neutral standard.
Agentic mode: coreAI builds its own context when the answer is missing
When coreAI cannot find the answer in the context it has already been given, it fetches it itself – in agentic mode the assistant combines semantic search, keyword search and filters against your own knowledge base, and calls external MCP tools if you have connected relevant sources. The result is more complete answers to multi-part questions, with source citations and an honest "I couldn't find it" when the answer genuinely isn't there.
The MCP server turns coreAI into a shared retrieval layer for AI agents
coreAI can expose the assistant's knowledge base as an MCP server, so external AI agents — developer tools, IDEs, and internal workflows — can ask the same questions as the chat via the standard Model Context Protocol. The same curated sources sit behind both the chat widget and the MCP surface, with no copying, no separate vector index, and no scraping.
The MCP import in coreAI is a client integration, not a bulk import
The MCP import connects coreAI to an external MCP server, lists which tools the server offers, and stores the tool schemas as JSON configuration on the assistant. The actual data is only fetched when the assistant needs it during a conversation – nothing is pushed into the knowledge base.
How coreAI uses entity types to distinguish products, documents, and contacts
coreAI indexes your data as typed entities — `products`, `contents`, `documents`, `events`, `educations`, `job_postings`, and `contacts` — so the assistant knows whether the answer should be about a product, a PDF, or a contact person. Each entity can also carry `properties` that are later used as filters in chat and search calls.
Protect an assistant with your own login via callbackUrl
Set a `callbackUrl` on the assistant and coreAI calls your own URL with the user's token when the widget initializes — and only lets the user through if the endpoint responds 200. The check runs once per widget load, not per question, and the login stays with you.