Jina AI MCP Server

Instructions

Implementation Reference

• index.ts:89-108 (handler)

  The groundStatement function executes the fact_check tool logic by querying Jina AI's grounding API with the provided statement and optional deepdive flag, parsing the response with GroundingResponseSchema.

  async function groundStatement(params: z.infer<typeof GroundingSchema>) {
    const headers: Record<string, string> = {
      'Authorization': `Bearer ${JINA_API_KEY}`,
      'Accept': 'application/json'
    };
  
    const statementQuery = encodeURIComponent(params.statement);
    const url = `https://g.jina.ai/${statementQuery}${params.deepdive ? '?deepdive=true' : ''}`;
  
    const response = await fetch(url, {
      method: 'GET',
      headers,
    });
  
    if (!response.ok) {
      throw new Error(`Jina AI Grounding API error: ${response.statusText}`);
    }
  
    return GroundingResponseSchema.parse(await response.json());
  }

• schemas.ts:68-71 (schema)

  Zod schema defining the input parameters for the fact_check tool: a required 'statement' string and optional 'deepdive' boolean.

  export const GroundingSchema = z.object({
    statement: z.string(),
    deepdive: z.boolean().optional().default(false)
  });

• index.ts:123-127 (registration)

  Registers the fact_check tool in the ListTools response, specifying its name, description, and input schema.

  {
    name: "fact_check",
    description: "Fact-check a statement using Jina AI's grounding engine",
    inputSchema: zodToJsonSchema(GroundingSchema)
  }

• index.ts:151-155 (handler)

  Switch case in CallToolRequestSchema handler that parses arguments for fact_check, calls groundStatement, and formats the response as text content.

  case "fact_check": {
    const args = GroundingSchema.parse(request.params.arguments);
    const result = await groundStatement(args);
    return { content: [{ type: "text", text: JSON.stringify(result, null, 2) }] };
  }

• evals.ts:25-32 (helper)

  Evaluation function (fact_checkEval) for testing the fact_check tool using mcp-evals framework.

  const fact_checkEval: EvalFunction = {
      name: 'fact_check Tool Evaluation',
      description: 'Evaluates the correctness of the fact-checking tool',
      run: async () => {
          const result = await grade(openai("gpt-4o"), "Is it true that the Great Wall of China is visible from space?");
          return JSON.parse(result);
      }
  };