Ultra MCP

Instructions

Implementation Reference

• src/server.ts:344-352 (registration)

  Registers the 'consensus' tool with the MCP server, providing title, description, input schema, and linking to the handler execution.
  // Register consensus tool server.registerTool("consensus", { title: "Consensus", description: "Get consensus from multiple AI models on a proposal", inputSchema: ConsensusSchema.shape, }, async (args) => { const aiHandlers = await getHandlers(); return await aiHandlers.handleConsensus(args); });
• src/server.ts:73-81 (schema)

  Zod schema defining the input structure for the consensus tool, including proposal, list of models with stances, and optional files.
  const ConsensusSchema = z.object({ proposal: z.string().describe("The proposal, idea, or decision to analyze from multiple perspectives"), models: z.array(z.object({ model: z.string().describe("Model name to consult (e.g., 'gemini-pro', 'gpt-4', 'gpt-5')"), stance: z.enum(["for", "against", "neutral"]).default("neutral").describe("Perspective stance for this model"), provider: z.enum(["openai", "gemini", "azure", "grok"]).optional().describe("AI provider for this model") })).min(1).describe("List of models to consult with their stances"), files: z.array(z.string()).optional().describe("Relevant file paths for context (optional)"), });
• src/handlers/ai-tools.ts:593-716 (handler)

  Core handler implementation that executes the consensus tool. Consults multiple AI models with configurable stances (for/against/neutral), collects individual analyses on the proposal, synthesizes a consensus summary, and returns structured results including all perspectives and recommendations.
  async handleConsensus(params: z.infer<typeof ConsensusSchema>) { const responses: any[] = []; // Consult each model with their specified stance for (const modelConfig of params.models) { const providerName = modelConfig.provider || (await this.providerManager.getPreferredProvider(['openai', 'gemini', 'azure', 'grok'])); const provider = await this.providerManager.getProvider(providerName); // Build stance-specific system prompt let stancePrompt = ""; switch (modelConfig.stance) { case "for": stancePrompt = "You are analyzing this proposal from a supportive perspective. Focus on benefits, opportunities, and positive aspects while being realistic about implementation."; break; case "against": stancePrompt = "You are analyzing this proposal from a critical perspective. Focus on risks, challenges, drawbacks, and potential issues while being fair and constructive."; break; case "neutral": default: stancePrompt = "You are analyzing this proposal from a balanced, neutral perspective. Consider both benefits and risks, opportunities and challenges equally."; break; } const systemPrompt = `${stancePrompt} Provide a thorough analysis of the proposal considering: - Technical feasibility and implementation complexity - Benefits and value proposition - Risks and potential challenges - Resource requirements and timeline considerations - Alternative approaches or modifications Be specific and actionable in your analysis.`; let prompt = `Analyze this proposal: ${params.proposal}`; if (params.files) { prompt += `\n\nRelevant files for context: ${params.files.join(", ")}`; } try { const response = await provider.generateText({ prompt, model: modelConfig.model, systemPrompt, temperature: 0.3, // Lower temperature for more consistent analysis useSearchGrounding: providerName === "gemini", toolName: 'consensus', }); responses.push({ model: modelConfig.model, provider: providerName, stance: modelConfig.stance, analysis: response.text, usage: response.usage, }); } catch (error) { responses.push({ model: modelConfig.model, provider: providerName, stance: modelConfig.stance, error: error instanceof Error ? error.message : "Unknown error", }); } } // Generate synthesis const synthesisPrompt = `Based on the following analyses from different perspectives, provide a comprehensive consensus summary: ${responses.map((r, i) => r.error ? `${i + 1}. ${r.model} (${r.stance}, ERROR): ${r.error}` : `${i + 1}. ${r.model} (${r.stance}): ${r.analysis}` ).join('\n\n')} Please synthesize these perspectives into: 1. **Key Points of Agreement**: What do most analyses agree on? 2. **Major Concerns and Disagreements**: Where do the analyses differ? 3. **Balanced Recommendation**: Based on all perspectives, what would you recommend? 4. **Next Steps**: What additional considerations or actions might be needed? Be objective and highlight both the strongest arguments for and against the proposal.`; const synthesisProvider = await this.providerManager.getProvider(await this.providerManager.getPreferredProvider(['openai', 'gemini', 'azure', 'grok'])); const synthesis = await synthesisProvider.generateText({ prompt: synthesisPrompt, systemPrompt: "You are an expert facilitator synthesizing multiple expert opinions. Provide balanced, objective analysis that captures the full spectrum of perspectives.", temperature: 0.4, useSearchGrounding: false, toolName: 'consensus', }); const result = { proposal: params.proposal, individual_analyses: responses, synthesis: synthesis.text, total_models_consulted: responses.length, successful_consultations: responses.filter(r => !r.error).length, }; return { content: [ { type: "text", text: JSON.stringify(result, null, 2), }, ], metadata: { toolName: "consensus", modelsConsulted: responses.length, synthesisModel: synthesis.model, totalUsage: responses.reduce((acc, r) => { if (r.usage) { return { inputTokens: (acc.inputTokens || 0) + (r.usage.inputTokens || 0), outputTokens: (acc.outputTokens || 0) + (r.usage.outputTokens || 0), totalTokens: (acc.totalTokens || 0) + (r.usage.totalTokens || 0), }; } return acc; }, {}), }, }; }
• src/handlers/ai-tools.ts:76-84 (schema)

  Zod schema for consensus tool inputs, used internally by the handler for type inference.
  const ConsensusSchema = z.object({ proposal: z.string().describe("The proposal, idea, or decision to analyze from multiple perspectives"), models: z.array(z.object({ model: z.string().describe("Model name to consult (e.g., 'gemini-pro', 'gpt-4', 'gpt-5')"), stance: z.enum(["for", "against", "neutral"]).default("neutral").describe("Perspective stance for this model"), provider: z.enum(["openai", "gemini", "azure", "grok"]).optional().describe("AI provider for this model") })).min(1).describe("List of models to consult with their stances"), files: z.array(z.string()).optional().describe("Relevant file paths for context (optional)"), });
• src/server.ts:701-719 (registration)

  Registers a prompt for the consensus tool, enabling natural language invocation with simplified args schema.
  server.registerPrompt("consensus", { title: "Multi-Model Consensus", description: "Get consensus from multiple AI models on a proposal or decision", // Use string-only schema for MCP compatibility argsSchema: { proposal: z.string().optional(), models: z.string().optional(), // e.g., "gpt-4:for, gemini:neutral" files: z.string().optional(), provider: z.string().optional(), }, }, (args) => ({ messages: [{ role: "user", content: { type: "text", text: `Get multi-model consensus on this proposal: ${args.proposal || 'Please provide a proposal or decision to analyze.'}\n\nConsult these models: ${args.models || 'gpt-4:neutral, gemini:neutral'}${args.files ? `\n\nRelevant files for context: ${args.files}` : ''}` } }] }));