protocol

mcp_tool_protocol

Protocol for Logtalk objects that provide tools to be exposed via an MCP (Model Context Protocol) server. Implements the MCP 2025-06-18 specification. Implementing objects must define the set of tools available and handle tool calls. Tool metadata (names, titles, descriptions, parameter schemas) can be derived automatically from info/2 and mode/2 directives on the tool predicates, or can be specified explicitly via the tool/1 predicate.

Availability:
logtalk_load(mcp_server(loader))
Author: Paulo Moura
Version: 0:3:0
Date: 2026-02-24
Compilation flags:
static
Dependencies:
(none)
Remarks:

  • Capabilities: Objects can optionally define capabilities/1 to declare which MCP capabilities they require. Currently supported: elicitation (allows the server to ask the user questions during tool execution via MCP elicitation). If capabilities/1 is not defined, only the tools capability is advertised.

  • Tool title: The tool title (human-friendly display name) is derived from a title key in the predicate info/2 directive. If not specified, the predicate name (functor) is used as the title.

  • Structured output: Tools can declare an output schema via output_schema/2 and return structured(StructuredContent) or structured(ContentItems, StructuredContent) results. The structured content is included in the structuredContent field of the tool call response alongside the content array.

  • Resource links: Tool results can include resource_link(URI, Name) or resource_link(URI, Name, Description, MimeType) content items in results/1 lists.

  • Elicitation: Tools that need user interaction during execution should define tool_call/4 instead of tool_call/3. The extra argument is an elicitation closure that can be called as call(Elicit, Message, Schema, Answer) where Message is the prompt text (an atom), Schema is a JSON Schema curly-term for the requested input, and Answer is unified with accept(Content), decline, or cancel.

Inherited public predicates:
(none)

Public predicates

capabilities/1

Returns a list of MCP capabilities required by this tool provider. Currently supported capabilities: elicitation (server can ask the user questions during tool execution). If not defined, only the tools capability is advertised. The server uses this to build the capabilities field in the initialize response.

Compilation flags:
static
Template:
capabilities(Capabilities)
Mode and number of proofs:
capabilities(-list(atom)) - one

tools/1

Returns a list of tool descriptors available from this object. Each descriptor is a compound term tool(Name, Functor, Arity) where Name is the MCP tool name (an atom), Functor is the predicate functor, and Arity is the predicate arity. By default, the tool name is the predicate functor. Tool titles are derived from info/2 title keys (falling back to the predicate name). Tool descriptions and parameter schemas are derived from the info/2 and mode/2 directives of the corresponding predicates.

Compilation flags:
static
Template:
tools(Tools)
Mode and number of proofs:
tools(-list(compound)) - one

tool_call/3

Handles a tool call. Name is the MCP tool name (as declared in tools/1), Arguments is a list of ArgumentName-Value pairs, and Result is unified with the tool result. The result must be one of: text(Atom) for a text result, error(Atom) for an error result, results(List) for a list of content items where each item is text(Atom), error(Atom), resource_link(URI, Name), or resource_link(URI, Name, Description, MimeType), structured(StructuredContent) for structured output with auto-generated text, or structured(ContentItems, StructuredContent) for structured output with explicit content items. If this predicate is not defined for a tool, the MCP server will use auto-dispatch: it calls the tool predicate as a message to the implementing object, collects output arguments, and returns them as a text result.

Compilation flags:
static
Template:
tool_call(Name,Arguments,Result)
Mode and number of proofs:
tool_call(+atom,+list(pair),--compound) - one

tool_call/4

Handles a tool call with elicitation support. Same as tool_call/3 but receives an Elicit closure as the third argument. The closure can be called as call(Elicit, Message, Schema, Answer) where Message is a prompt text (an atom), Schema is a curly-term JSON Schema describing the requested input (e.g., {type-object, properties-{answer-{type-string, enum-[yes, no]}}, required-[answer]}), and Answer is unified with accept(Content) (where Content is the user response as a curly-term), decline, or cancel. Only available when the application declares elicitation in its capabilities/1. Falls back to tool_call/3 and then auto-dispatch if not defined.

Compilation flags:
static
Template:
tool_call(Name,Arguments,Elicit,Result)
Mode and number of proofs:
tool_call(+atom,+list(pair),+callable,--compound) - one

output_schema/2

Returns the JSON output schema for the given tool name. Optional; when defined, the schema is included in the tool descriptor as outputSchema and the tool can return structured(StructuredContent) or structured(ContentItems, StructuredContent) results. The schema must be a curly-term following JSON Schema format (e.g. {type-object, properties-{temperature-{type-number}}, required-[temperature]}).

Compilation flags:
static
Template:
output_schema(Name,Schema)
Mode and number of proofs:
output_schema(+atom,-compound) - zero_or_one

Protected predicates

(none)

Private predicates

(none)

Operators

(none)