Build a MiniMax H3 Omni-Modal Media MCP Server for Agent-Driven Video & Audio Generation in 2026

MiniMax H3 (open weights August 3, 2026) is the first fully open-source omni-modal model—a 33B-parameter system that understands text, image, and video while generating 4-15 second 2K video clips with native stereo audio. This FastMCP TypeScript server exposes H3's generation capabilities as MCP tools, letting AI agents in Claude Desktop and Cursor produce multi-modal content without manual prompt engineering.

Server Architecture

The server implements 5 MCP tools: generate_video (text/image-to-video), generate_audio (text-to-speech with stereo), analyze_media (understand existing video/image), extract_frames (keyframe extraction), and compose_scene (multi-step scene composition). Each tool handles the full pipeline from prompt to final media file.

// minimax-h3-mcp/server.ts
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";
import axios from "axios";
import * as fs from "fs/promises";

const server = new McpServer({ name: "minimax-h3-media", version: "1.0.0" });
const H3_API = process.env.MINIMAX_H3_API_URL || "https://api.minimax.chat/v1";

// Tool 1: Generate Video
server.tool(
  "generate_video",
  "Generate 2K video from text prompt or reference image",
  {
    prompt: z.string().describe("Text description of the video scene"),
    reference_image: z.string().optional().describe("Base64 reference image for img2vid"),
    duration: z.number().min(4).max(15).default(5).describe("Duration in seconds"),
    resolution: z.enum(["720p", "1080p", "2k"]).default("1080p")
  },
  async ({ prompt, reference_image, duration, resolution }) => {
    const response = await axios.post(`${H3_API}/video/generate`, {
      prompt,
      image: reference_image ? `data:image/jpeg;base64,${reference_image}` : undefined,
      duration_seconds: duration,
      resolution,
      audio: { enabled: true, stereo: true }
    }, {
      headers: { "Authorization": `Bearer ${process.env.MINIMAX_H3_API_KEY}` }
    });
    
    const videoUrl = response.data.video_url;
    const outputPath = `/tmp/h3_video_${Date.now()}.mp4`;
    const videoData = await axios.get(videoUrl, { responseType: "arraybuffer" });
    await fs.writeFile(outputPath, videoData.data);
    
    return { content: [{ type: "text", text: `Video generated: ${outputPath}
Duration: ${duration}s
Resolution: ${resolution}
Audio: stereo` }] };
  }
);

// Tool 2: Generate Audio
server.tool(
  "generate_audio",
  "Generate stereo audio from text with voice cloning support",
  {
    text: z.string().describe("Text to synthesize"),
    voice_id: z.string().optional().describe("Voice clone ID"),
    language: z.enum(["en", "es", "fr", "de", "ja", "ko"]).default("en")
  },
  async ({ text, voice_id, language }) => {
    const response = await axios.post(`${H3_API}/audio/generate`, {
      text, voice_id, language, stereo: true,
      sample_rate: 44100, format: "wav"
    }, {
      headers: { "Authorization": `Bearer ${process.env.MINIMAX_H3_API_KEY}` }
    });
    const outputPath = `/tmp/h3_audio_${Date.now()}.wav`;
    const audioData = await axios.get(response.data.audio_url, { responseType: "arraybuffer" });
    await fs.writeFile(outputPath, audioData.data);
    return { content: [{ type: "text", text: `Audio generated: ${outputPath}` }] };
  }
);

// Tool 3: Analyze Media
server.tool(
  "analyze_media",
  "Analyze video or image content for scene description and metadata",
  {
    media_path: z.string().describe("Path to video or image file"),
    analysis_type: z.enum(["describe", "ocr", "objects", "sentiment"]).default("describe")
  },
  async ({ media_path, analysis_type }) => {
    const mediaBuffer = await fs.readFile(media_path);
    const base64 = mediaBuffer.toString("base64");
    const response = await axios.post(`${H3_API}/analyze`, {
      media: `data:${media_path.endsWith(".mp4") ? "video" : "image"}/base64,${base64}`,
      analysis_type
    }, {
      headers: { "Authorization": `Bearer ${process.env.MINIMAX_H3_API_KEY}` }
    });
    return { content: [{ type: "text", text: JSON.stringify(response.data.result, null, 2) }] };
  }
);

server.connect();

Media Generation Cost Table

Production Reality Check

MiniMax H3 has a license caveat: it excludes US/EU commercial use under the MiniMax license. For commercial deployments in those regions, verify licensing terms or use the MiniMax API (which has separate commercial terms). The open weights are suitable for research, evaluation, and non-commercial deployments. At SaaSNext, we use H3 for internal prototyping and content ideation, routing production content generation to commercially licensed alternatives.

For related media generation patterns, see the Veo 3.1 & Seedream 5.0 Media MCP Server. The MCP Directory has complementary tools for multi-modal agent pipelines.

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last tested: August 2026 with Node v22, FastMCP 4.0.0b3, MiniMax H3 33B, and MCP SDK 2026-07-28.