ConFuse

    Using AI Agents with ConFuse

    You can connect an AI assistant — a coding agent, a chat assistant, or your own tooling — to ConFuse so that it answers questions using your knowledge base rather than its own general knowledge.

    What this changes

    Without a connection, an assistant answers from what it learned during training. That is often reasonable and sometimes badly wrong about your project.

    With a connection, the assistant can look things up in your material before answering. The practical difference:

    Unconnected assistantConnected to ConFuse
    Source of truthGeneral knowledgeYour connected sources
    About your projectOften generic or wrongSpecific and current
    EvidenceUsually noneCited
    Unknown topicsConfidently guessesSays it is not covered

    How you connect

    ConFuse publishes an endpoint that assistants supporting the Model Context Protocol (MCP) can connect to. MCP is an open standard that lets AI tools use external data sources.

    You will need:

    1. Access to your ConFuse workspace.
    2. An assistant that supports MCP connections.
    3. The connection details for your workspace — ask your workspace administrator if you do not have them.

    The exact setup differs per assistant; it is normally a matter of adding a server endpoint and signing in.

    What the assistant can do

    • Search your connected sources to answer questions.
    • Cite the material behind its claims.
    • Report when the sources do not cover a question.
    • Assemble context from several places to answer a cross-cutting question.

    What it should not do

    A well-behaved assistant connected to ConFuse will not:

    • Answer project questions from general knowledge when your sources have the answer.
    • Present uncited claims as project fact.
    • Hide gaps. If the material is missing, it should say so.
    • Invent file names, figures or quotes.

    If yours does these things, the material is there to correct it — check the citations and ask again with the missing piece named.

    Getting the best results

    • Give it the context you would give a colleague. "In the payments service, how are chargebacks handled?" beats "how do chargebacks work?".
    • Ask for citations. It forces the assistant to show its working.
    • Ask it to flag uncertainty. "Answer, then list anything you could not verify" is a very effective prompt pattern.
    • Ask it to verify. For important answers, ask the assistant to check each claim against the sources before presenting it.

    Safety and boundaries

    • The assistant only sees material the workspace is permitted to read.
    • Connected sources are not shared outside your workspace.
    • Nothing about how ConFuse is built is exposed to the assistant — it sees your knowledge, not our internals.
    • Humans stay in control: connected material is for answering questions, not for taking actions in your systems.

    See Security and Privacy for the full picture.

    Related reading