MCP Gemini Server

import { z } from "zod"; import { ModelNameSchema, ModelPreferencesSchema, } from "./schemas/CommonSchemas.js"; // Tool Name export const GEMINI_CHAT_TOOL_NAME = "gemini_chat"; // Tool Description export const GEMINI_CHAT_TOOL_DESCRIPTION = ` Manages stateful chat sessions with Google Gemini models. This consolidated tool supports three operations: - start: Initiates a new chat session with optional history and configuration - send_message: Sends a text message to an existing chat session - send_function_result: Sends function execution results back to a chat session Each operation returns appropriate responses including session IDs, model responses, or function call requests. `; // Operation enum for chat actions export const chatOperationSchema = z .enum(["start", "send_message", "send_function_result"]) .describe("The chat operation to perform"); // Zod Schema for thinking configuration (reused from content generation) export const thinkingConfigSchema = z .object({ thinkingBudget: z .number() .int() .min(0) .max(24576) .optional() .describe( "Controls the amount of reasoning the model performs. Range: 0-24576. Lower values provide faster responses, higher values improve complex reasoning." ), reasoningEffort: z .enum(["none", "low", "medium", "high"]) .optional() .describe( "Simplified control over model reasoning. Options: none (0 tokens), low (1K tokens), medium (8K tokens), high (24K tokens)." ), }) .optional() .describe("Optional configuration for controlling model reasoning."); // Generation config schema export const generationConfigSchema = z .object({ temperature: z .number() .min(0) .max(1) .optional() .describe( "Controls randomness. Lower values (~0.2) make output more deterministic, higher values (~0.8) make it more creative. Default varies by model." ), topP: z .number() .min(0) .max(1) .optional() .describe( "Nucleus sampling parameter. The model considers only tokens with probability mass summing to this value. Default varies by model." ), topK: z .number() .int() .min(1) .optional() .describe( "Top-k sampling parameter. The model considers the k most probable tokens. Default varies by model." ), maxOutputTokens: z .number() .int() .min(1) .optional() .describe("Maximum number of tokens to generate in the response."), stopSequences: z .array(z.string()) .optional() .describe("Sequences where the API will stop generating further tokens."), thinkingConfig: thinkingConfigSchema, }) .optional() .describe("Optional configuration for controlling the generation process."); // Safety setting schemas export const harmCategorySchema = z .enum([ "HARM_CATEGORY_UNSPECIFIED", "HARM_CATEGORY_HATE_SPEECH", "HARM_CATEGORY_SEXUALLY_EXPLICIT", "HARM_CATEGORY_HARASSMENT", "HARM_CATEGORY_DANGEROUS_CONTENT", ]) .describe("Category of harmful content to apply safety settings for."); export const harmBlockThresholdSchema = z .enum([ "HARM_BLOCK_THRESHOLD_UNSPECIFIED", "BLOCK_LOW_AND_ABOVE", "BLOCK_MEDIUM_AND_ABOVE", "BLOCK_ONLY_HIGH", "BLOCK_NONE", ]) .describe( "Threshold for blocking harmful content. Higher thresholds block more content." ); export const safetySettingSchema = z .object({ category: harmCategorySchema, threshold: harmBlockThresholdSchema, }) .describe( "Setting for controlling content safety for a specific harm category." ); // History schemas for chat initialization const historyPartSchema = z .object({ text: z.string().describe("Text content of the part."), // Note: Could add other part types like inlineData, functionCall, functionResponse later }) .describe( "A part of a historical message, primarily text for initialization." ); const historyContentSchema = z .object({ role: z .enum(["user", "model"]) .describe( "The role of the entity that generated this content (user or model)." ), parts: z .array(historyPartSchema) .min(1) .describe("An array of Parts making up the message content."), }) .describe("A single message turn in the conversation history."); // Function declaration schemas (for tools) const functionParameterTypeSchema = z .enum(["OBJECT", "STRING", "NUMBER", "BOOLEAN", "ARRAY", "INTEGER"]) .describe("The data type of the function parameter."); const baseFunctionParameterSchema = z.object({ type: functionParameterTypeSchema, description: z .string() .optional() .describe("Description of the parameter's purpose."), enum: z .array(z.string()) .optional() .describe("Allowed string values for an ENUM-like parameter."), }); type FunctionParameterSchemaType = z.infer< typeof baseFunctionParameterSchema > & { properties?: { [key: string]: FunctionParameterSchemaType }; required?: string[]; items?: FunctionParameterSchemaType; }; const functionParameterSchema: z.ZodType<FunctionParameterSchemaType> = baseFunctionParameterSchema .extend({ properties: z.lazy(() => z.record(functionParameterSchema).optional()), required: z.lazy(() => z .array(z.string()) .optional() .describe("List of required property names for OBJECT types.") ), items: z.lazy(() => functionParameterSchema .optional() .describe( "Defines the schema for items if the parameter type is ARRAY." ) ), }) .describe( "Schema defining a single parameter for a function declaration, potentially recursive." ); const functionParameterPropertiesSchema = z .record(functionParameterSchema) .describe("Defines nested properties if the parameter type is OBJECT."); export const functionDeclarationSchema = z .object({ name: z.string().min(1).describe("The name of the function to be called."), description: z .string() .min(1) .describe("A description of what the function does."), parameters: z .object({ type: z .literal("OBJECT") .describe("The top-level parameters structure must be an OBJECT."), properties: functionParameterPropertiesSchema.describe( "Defines the parameters the function accepts." ), required: z .array(z.string()) .optional() .describe("List of required parameter names at the top level."), }) .describe("Schema defining the parameters the function accepts."), }) .describe( "Declaration of a single function that the Gemini model can request to call." ); // Tool schema for chat const toolSchema = z .object({ functionDeclarations: z .array(functionDeclarationSchema) .optional() .describe("List of function declarations for this tool."), }) .describe("Represents a tool definition containing function declarations."); // Tool config schema const functionCallingConfigSchema = z .object({ mode: z .enum(["AUTO", "ANY", "NONE"]) .optional() .describe("The function calling mode."), allowedFunctionNames: z .array(z.string()) .optional() .describe("Optional list of function names allowed."), }) .optional(); const toolConfigSchema = z .object({ functionCallingConfig: functionCallingConfigSchema, }) .optional() .describe( "Optional. Per-request tool configuration, e.g., to force function calling mode." ); // Function response schema for send_function_result const functionResponseInputSchema = z .object({ name: z .string() .min(1) .describe( "Required. The name of the function that was called by the model." ), response: z .record(z.unknown()) .describe( "Required. The JSON object result returned by the function execution." ), }) .describe( "Represents the result of a single function execution to be sent back to the model." ); // System instruction schema const systemInstructionSchema = z .object({ parts: z.array( z.object({ text: z.string(), }) ), }) .optional() .describe( "Optional. A system instruction to guide the model's behavior for the entire session." ); // Main parameters schema with conditional fields based on operation export const GEMINI_CHAT_PARAMS = { operation: chatOperationSchema, // Common fields used across operations sessionId: z .string() .uuid() .optional() .describe( "Required for 'send_message' and 'send_function_result'. The unique identifier of the chat session." ), // Fields for 'start' operation modelName: ModelNameSchema.optional().describe( "Optional for 'start'. The name of the Gemini model to use for this chat session. If omitted, uses server default." ), history: z .array(historyContentSchema) .optional() .describe( "Optional for 'start'. Initial conversation turns to seed the chat session. Must alternate between 'user' and 'model' roles." ), systemInstruction: systemInstructionSchema, // Fields for 'send_message' operation message: z .string() .min(1) .optional() .describe( "Required for 'send_message'. The text message content to send to the model." ), // Fields for 'send_function_result' operation functionResponses: z .array(functionResponseInputSchema) .min(1) .optional() .describe( "Required for 'send_function_result'. Array containing the results of function calls executed by the client. Note: This array is JSON.stringify'd before being passed to the Gemini API." ), // Shared optional configuration fields tools: z .array(toolSchema) .optional() .describe( "Optional. Tools (function declarations) the model may use. For 'start', sets session-wide tools. For 'send_message', overrides session tools for this turn." ), toolConfig: toolConfigSchema, generationConfig: generationConfigSchema, safetySettings: z .array(safetySettingSchema) .optional() .describe( "Optional. Safety settings to apply. For 'start', sets session-wide settings. For other operations, overrides session settings for this turn." ), cachedContentName: z .string() .min(1) .optional() .describe( "Optional. Identifier for cached content in format 'cachedContents/...' to use with this operation." ), modelPreferences: ModelPreferencesSchema.optional(), }; // Type helper export type GeminiChatArgs = z.infer<z.ZodObject<typeof GEMINI_CHAT_PARAMS>>;