Perplexity Agent MCP

Instructions

Implementation Reference

• src/tools/answer.ts:25-35 (handler)

  The handler function for the 'answer' tool. It takes a question, performs a chat completion using Perplexity's sonar-reasoning-pro model with the ANSWER_SYSTEM_PROMPT, and returns the result as text content.

  async ({ question }) => {
    const result = await performChatCompletion(
      [{ role: "user", content: question }],
      {
        model: "sonar-reasoning-pro",
        system: ANSWER_SYSTEM_PROMPT,
        searchContextSize: "high",
      },
    );
    return { content: [{ type: "text", text: result }] };
  },

• src/tools/answer.ts:13-24 (schema)

  Tool schema definition including description and inputSchema with 'question' parameter using Zod validation.

      {
        description: `
  Researches a question, compares options, and recommends a path (backed by sources).
  Use for library choices, architecture trade-offs, migrations, complex debugging, and performance decisions.
  Returns a concise recommendation, a brief why, and short how-to steps.
  Examples: "Should I use Zod or Valibot?", "How to optimize React bundle size?", "Best auth approach for Node.js microservices?"
  One question per call—split combined requests into separate queries.
  `.trim(),
        inputSchema: {
          question: z.string().describe("The decision or problem to answer"),
        },
      },

• src/tools/answer.ts:11-36 (registration)

  The server.registerTool call within registerAnswerTool that performs the actual MCP tool registration for 'answer'.

    server.registerTool(
      "answer",
      {
        description: `
  Researches a question, compares options, and recommends a path (backed by sources).
  Use for library choices, architecture trade-offs, migrations, complex debugging, and performance decisions.
  Returns a concise recommendation, a brief why, and short how-to steps.
  Examples: "Should I use Zod or Valibot?", "How to optimize React bundle size?", "Best auth approach for Node.js microservices?"
  One question per call—split combined requests into separate queries.
  `.trim(),
        inputSchema: {
          question: z.string().describe("The decision or problem to answer"),
        },
      },
      async ({ question }) => {
        const result = await performChatCompletion(
          [{ role: "user", content: question }],
          {
            model: "sonar-reasoning-pro",
            system: ANSWER_SYSTEM_PROMPT,
            searchContextSize: "high",
          },
        );
        return { content: [{ type: "text", text: result }] };
      },
    );

• src/server.ts:17-18 (registration)

  Invocation of registerAnswerTool during MCP server creation to register the 'answer' tool.

  registerLookupTool(server);
  registerAnswerTool(server);

• src/prompts.ts:73-102 (helper)

  The system prompt used in the 'answer' tool handler for guiding the AI in technical decision making.

  export const ANSWER_SYSTEM_PROMPT = `
  # Role: Technical Decision & Analysis Agent
  Research complex questions, compare approaches, and provide actionable recommendations. Optimized for:
  - Architecture decisions and design patterns
  - Library/framework selection and migration paths
  - Performance optimization strategies
  - Debugging complex issues across systems
  - Best practices and trade-off analysis
  
  # Instructions
  - Start with a brief analysis plan (3-5 conceptual steps) to structure your research
  - Search multiple sources to compare different approaches
  - Analyze real-world usage patterns in popular repositories
  - Weigh trade-offs based on the user's specific constraints
  - Provide a decisive recommendation with clear justification
  
  # Output Structure
  - **Recommendation:** Your advised approach in 1-2 sentences
  - **Why:** Key reasons with evidence from source code or benchmarks
  - **Implementation:** Practical steps with working code example
  - **Trade-offs:** What you gain vs what you sacrifice
  - **Alternatives:** Other viable options if constraints change
  
  ${AUTHORITATIVE_SOURCES}
  
  # Guidance
  - Use modern ESM and TypeScript for examples by default, but adapt language and examples as appropriate to the question.
  - Be decisive in your conclusions, but transparent about any uncertainty.
  - Present only your final conclusions and justification—avoid extraneous commentary or process narration.
  `.trim();