Knowledge · MCP

    How AI Discovers MCP Capabilities

    AI does not need to be trained on your software. It reads the tool catalog at connection time and selects actions from descriptions and schemas.

    How does an AI know which tools it can use?

    At connection time the AI client asks the server for its tool list and receives names, descriptions and input schemas. It selects a tool by matching the user's intent to those descriptions, fills the arguments from context, and calls it. No training, no hard-coded integration, no redeployment when tools change.

    Key takeaways

    • Discovery happens at runtime, not at training time.
    • The description is what the model reasons over — precision matters.
    • New tools become usable the moment they are published.
    • Ambiguous or overlapping descriptions cause wrong selection.
    • Schemas keep the model from inventing arguments.

    What discovery replaces

    In a traditional integration, someone writes code for each supported action. Adding a capability means a development cycle on both sides.

    With discovery the capability list is data. Publish a new tool and the agent can use it in the next conversation.

    The selection process, step by step

    • The client connects and requests the tool catalog.
    • The model reads names, descriptions and schemas.
    • The user's request is matched against those descriptions.
    • Arguments are extracted from the conversation and prior tool results.
    • The call is made; the result feeds the next decision.

    Why descriptions decide accuracy

    The model has no other information about your system. Two tools described as 'update a record' are indistinguishable; 'reschedule an existing appointment to a new time slot' and 'change the assigned technician on an appointment' are not.

    Stating what a tool must not be used for is as valuable as stating what it does.

    Chaining discovered tools

    Real requests span several tools: find the customer, list their open jobs, create an invoice, send it. The output of one call becomes the input of the next.

    This is only reliable when tools return canonical identifiers rather than display names.

    What discovery does not grant

    Discovery reveals capability, not permission. A tool can appear in the catalog and still be refused at call time because the session lacks the required scope. Visibility and authorisation are separate layers.

    Where URBLD fits

    URBLD publishes a versioned tool catalog; when new operational actions ship, connected agents pick them up on reconnect without any integration work.

    FAQ

    Frequently Asked Questions

    Straight answers about how URBLD runs the business end-to-end.

    More in MCP

    The protocol that lets AI do work instead of describing it.

    Browse MCP
    Share this page