Fusion MCP ServerDocumentation

Menu

Prompts

A prompt is a template the user picks - a slash command in an IDE, an entry in a menu - and not something the model calls on its own. Flag a JSON-RPC method with @MCPPrompt:

@ApplicationScoped
public class MyPrompts {
    @MCPPrompt
    @JsonRpc(value = "my/review", documentation = "Reviews a code snippet.")
    public PromptResponse review(
            @JsonRpcParam(required = true, documentation = "The code to review.") final String code,
            @JsonRpcParam(documentation = "Language of the snippet.") final String language) {
        return new PromptResponse(
                null,
                "Code review",
                List.of(new PromptResponse.Message(
                        Role.user,
                        Content.text("Review this " + (language == null ? "code" : language) + ":\n" + code))));
    }
}

Two constraints, both coming from the specification:

  • the method must return a PromptResponse - or a CompletionStage<PromptResponse>,

  • prompt arguments are always strings, so only declare String parameters.

prompts/list exposes the name, the description and the arguments - name, description and whether it is required - all derived from the signature exactly like for a tool. prompts/get invokes the method and returns the expanded messages.

Completing the arguments

Implement MCPCompletions to suggest values while the user types an argument. Return null when the request is not for you: every implementation is visited, in bean priority order (@Order), until one answers.

@ApplicationScoped
public class LanguageCompletions implements MCPCompletions {
    private static final List<String> LANGUAGES = List.of("java", "javascript", "python", "rust");

    @Override
    public CompleteResult.Completion complete(final CompletionRef ref, final CompletionArgument argument,
                                              final CompletionContext context) {
        if (!"ref/prompt".equals(ref.type()) || !"my/review".equals(ref.name()) || !"language".equals(argument.name())) {
            return null;
        }
        return MCPCompletions.matching(LANGUAGES, argument.value());
    }
}

MCPCompletions.matching(candidates, prefix) filters on the prefix and truncates to the 100 values MCP allows, setting hasMore accordingly. Build the Completion yourself for anything smarter - a database lookup, a fuzzy match.

CompletionContext carries the arguments already resolved, which is what dependent completions need:

// suggest the cities of the country the user already picked
final var country = context == null || context.arguments() == null ? null : context.arguments().get("country");
return MCPCompletions.matching(cities(country), argument.value());

ref.type() is ref/prompt for a prompt argument and ref/resource for a resource template variable - in the latter case ref.uri() holds the template.

NOTE

the completions capability is only advertised when at least one MCPCompletions bean exists, so a client never offers an empty completion.

Navigation