Provides MCP tools to generate, edit, transcribe, upload, and query media through AgentScope Platform models.
list_models: list nine configured models, optionally filtered by capability.
synthesize_speech: generate speech with TTS models using required text and voice; optional format, instruction, and sample rate.
generate_image / edit_image: generate or edit images with qwen-image-3.0 using prompts and public image URLs.
transcribe_audio: transcribe a public audio URL with qwen-audio-3.0-asr-flash.
generate_video: submit text-to-video, image-to-video, or reference-to-video tasks; returns task_id for polling.
edit_video: submit HappyHorse video editing tasks with up to five reference images; returns task_id.
get_task: query an upstream task by task_id.
upload_media: upload base64 data URLs (default decoded limit 10 MiB) and return media URLs.
Media calls require AGENTSCOPE_API_KEY; list_models works without credentials or network access.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@AgentScope Media MCP Generate an image of a red fox in a snowy forest at sunrise."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Python MCP 服务：通过 AgentScope Platform 调用语音、图片、ASR 和视频模型。无网页、无本地模型、无账户 JWT 接口。
状态：开发版本。已开始真实接口联调，逐项结果见 验证记录。已发布到 PyPI，版本 0.1.0。
安装 uv 后，在项目目录运行：
uv sync --locked
uv run agentscope-media-mcpstdio 服务由 MCP 客户端启动；直接在终端运行时等待协议输入属于正常行为。
密钥通过启动进程的 AGENTSCOPE_API_KEY 环境变量提供，不要写进源码。
未填写密钥仍可以初始化和列出工具，调用媒体接口时会返回配置错误。
本地客户端配置：
{
"mcpServers": {
"agentscope-media": {
"command": "uv",
"args": ["--directory", "/home/fanxing/repos/agentscope-media-mcp", "run", "--locked", "agentscope-media-mcp"],
"env": {"AGENTSCOPE_API_KEY": "<填写模型API Key>"}
}
}
}Related MCP server: MiniMax MCP JS
工具 | 模型/功能 |
list_models | 九个模型的静态清单，可按能力筛选 |
synthesize_speech | qwen-audio-3.0-tts-flash / qwen-audio-3.0-tts-plus |
generate_image / edit_image | qwen-image-3.0 |
transcribe_audio | qwen-audio-3.0-asr-flash |
generate_video | happyhorse-1.1-t2v / i2v / r2v、wan3.0-video |
edit_video | happyhorse-1.0-video-edit |
upload_media | 将 base64 data URL 上传到平台 |
get_task | 根据上游 task_id 查询结果 |
TTS 必须指定该模型支持的 voice，避免猜测音色。Wan 文本输入已通过一次真实生成测试，其他输入模式尚未验证。
480P、720P、1080P 可作为视频分辨率输入；实际可用性以具体模型和网关验证为准。
图片尺寸使用 1024*1024 形式，不保证任意尺寸都被模型接受。
配音：选择 Flash 或 Plus，填写文本和有效 voice ID，调用 synthesize_speech。
图生视频：先取得公共图片 URL，再指定 happyhorse-1.1-i2v 和 image_url。
参考生视频：指定 happyhorse-1.1-r2v 和 reference_urls，可在提示词中引用 [Image 1]。
视频任务提交后保存 task_id，间隔查询 get_task，直到成功或失败。
图片、音频结果直接返回媒体 URL 或文本；如果上游返回任务，则保留任务编号。
服务不下载任意 URL，不读取调用者电脑上的路径，不永久保存作品。 API 返回链接可能过期，需要由调用者及时保存。 生成 POST 请求不自动重试；submission_unknown 表示受理结果不明，不应盲目重复生成。
环境变量 | 默认值 |
AGENTSCOPE_API_KEY | 无；生成请求必需 |
AGENTSCOPE_MEDIA_BASE_URL | |
AGENTSCOPE_HTTP_CONCURRENCY | 4 |
AGENTSCOPE_CONNECT_TIMEOUT | 15 秒 |
AGENTSCOPE_READ_TIMEOUT | 180 秒 |
AGENTSCOPE_UPLOAD_MAX_BYTES | 10485760（10 MiB） |
并发限制是同时进行的 HTTP 请求数，不是上游运行中的视频任务数。 不连接 Credits、定价和用户信息接口，无费用审批。 媒体 HTTP 客户端使用直连，不继承本机代理变量；不会修改系统或其他应用的网络配置。
uv sync --locked --group dev
uv run pytest
uv run ruff check .
uv build
uv run python scripts/check_wheel.py已发布到 PyPI。魔搭选择 uvx，固定已验证版本：
{
"mcpServers": {
"agentscope-media": {
"command": "uvx",
"args": ["agentscope-media-mcp==0.1.0"],
"env": {"AGENTSCOPE_API_KEY": "<部署时填写>"}
}
}
}魔搭托管运行 stdio 进程并提供远程连接，程序本身不监听 HTTP 端口。 平台的密钥填写和托管限制需实际部署验证。
请求字段参考及验证边界见 接口说明。
Edit reference images using qwen-image-3.0 and public image URLs.
Inputs: prompt (required), image_urls (required), seed, size. Returns: object
Submit HappyHorse video editing with up to five reference images; returns task_id.
Inputs: prompt (required), video_url (required), seed, resolution, +1 more. Returns: object
Generate an image with qwen-image-3.0. Optional size uses width*height syntax.
Inputs: prompt (required), seed, size, negative_prompt. Returns: object
Submit video generation; query task_id later. i2v needs image_url; r2v needs 1–9 reference_urls. Wan text-only mapping passed one smoke test; other modes are unverified. Omit optional parameters to use upstream defaults.
Inputs: model (required), prompt (required), seed, ratio, +4 more. Returns: object
Query an upstream task using the same API Key. Keep polling explicitly at intervals; pending/running is not generation success. UNKNOWN may mean the task expired.
Inputs: task_id (required). Returns: object
List all nine configured models; no credentials or network access required.
Inputs: capability. Returns: object
Synthesize speech. Supply a voice ID supported by the chosen model; none is guessed.
Inputs: text (required), voice (required), model, format, +2 more. Returns: object
Transcribe a public audio URL using qwen-audio-3.0-asr-flash.
Inputs: audio_url (required), language, enable_itn. Returns: object
Upload base64 image/audio/video data (default decoded limit 10 MiB). Returns media URLs. Prefer existing public URLs for large files; local paths are not accepted.
Inputs: data_url (required). Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 9 tools
Each tool targets a distinct resource+action (models, image gen/edit, video gen/edit, TTS, ASR, upload, task polling). The generate_image/edit_image and generate_video/edit_video pairs could superficially be confused, but descriptions clarify the difference (from-scratch vs reference-based).
All nine tools follow a consistent snake_case verb_noun pattern (list_models, generate_image, edit_video, synthesize_speech, transcribe_audio, get_task, upload_media). No mixed conventions.
Nine tools is well-scoped for a media generation/editing server, with one tool per major capability plus task polling and media upload. No redundancy or dead weight.
Surface covers the full media lifecycle: generation, editing, upload, transcription, TTS, and async task polling. Minor gaps remain — no way to discover valid voice IDs (required by synthesize_speech but 'none is guessed'), and no task cancel/list operations.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
AI video, image and audio generation with 140+ AI models from any MCP agent
Discover and call AI agents via MCP. Supports A2A agents and platform agents with async tasks.
Prepaid inference for agents over hosted MCP. Chat, image, and video.
