H3 Cloud MCP
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., "@H3 Cloud MCP Check the ComfyUI backend, then generate a video of rain on neon Taipei streets at night"
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.
H3 Cloud MCP
用本地 MCP 客户端调用 AutoDL Art 的 MiniMax H3 托管工作流:提交文生、首尾帧、图片或音频参考任务,查询状态并下载 MP4。视频由 AutoDL Art 托管生成,使用者无需为这条路线单独租 GPU 或下载模型权重。
**当前状态:**MCP 已按 AutoDL Art 的参数表接入目录中的 18 个 H3 工作流,覆盖文生、首尾帧、多图和多音频。文生视频链路真实跑通两次:两条 5 秒、480p 横屏样片分别实扣 ¥0.20,产物均为 864×480、5.167 秒的 H.264/AAC MP4。其他工作流已做参数校验与本地模拟测试,尚未逐个付费出片。仓库不含样片、令牌或个人资料。此 MCP 在你的电脑上以 stdio 运行,并不是已经部署好的公共在线服务。
完整使用说明书:从注册/登录、实名认证、设置限额令牌,到安装、出片、核对账单和常见问题。首次使用请按说明书操作。
能做什么
已接通的托管工具 | 用途 |
| 按 2026-10-26 观察价格本地估算,不调用平台,不产生费用。 |
| 提交一条付费文生视频任务,返回 |
| 列出 18 个工作流的类别、分辨率、时长和素材要求。 |
| 按所选工作流估算价格,不产生费用。 |
| 按工作流 ID 提交文生、首尾帧、多图或多音频任务;素材使用 HTTPS URL。 |
| 根据 |
| 下载完成的 MP4 到本机 |
原有的 h3_art_generate 是已实测文生工作流 minimax_h3_lightx2v_no_pic 的快捷入口;新工具按各工作流要求选择分辨率、时长、必填图片与音频。18 个工作流的价格已核对,但只有文生工作流真实出片验收。图片和音频输入当前使用可访问的 HTTPS URL;本机文件上传尚未接入。各类别及分辨率、调用示例。
Related MCP server: mcp-minimax
五步开始(macOS)
在 AutoDL Art 登录;若平台要求,先完成实名认证。在令牌管理创建 ComfyUI 分组令牌,设置你能接受的总额度与短有效期。创建后检查页面显示的实际额度。
安装 Python 3.11+、
uv;克隆仓库并安装依赖:git clone https://github.com/lst016/h3-cloud-mcp.git cd h3-cloud-mcp uv sync在你自己的交互式终端中,将令牌存入 macOS 钥匙串:
security add-generic-password -U -a "$(id -un)" -s h3-cloud-mcp-autodl-art -w命令会提示输入令牌;不要把令牌写进命令参数、MCP 配置或 Git。其他操作系统可用自己的秘密管理器向 MCP 进程提供
AUTODL_ART_TOKEN。在仓库目录运行下列命令,把脚本的绝对路径注册到 Codex CLI;其他 MCP 客户端配置为同一个 stdio 命令。
codex mcp add h3-art -- /absolute/path/h3-cloud-mcp/scripts/run-art-mcp-macos.sh打开所选工作流的价格页核对实时单价,再在 MCP 客户端调用
h3_art_list_workflows→h3_art_quote_workflow→h3_art_generate_workflow→h3_art_job_status→h3_art_download_video。只做已实测文生时,也可使用快捷工具:h3_art_generate({"prompt":"一只透明玻璃鲸鱼在桌面上方缓慢游动,单镜头,无文字","duration":5,"resolution":"480p横"})。
完整参数、Codex 配置、Windows/Linux 说明和出错处理见使用说明书。scripts/run-art-mcp-macos.sh 会启用付费提交,并在每个 MCP 进程中限制最多 1 条任务、按已观察高峰价估算最多 ¥0.25。这个本地估算不是平台账单上限;平台令牌额度要单独设置。提交超时后先查平台任务与账单,不要盲目重复提交。
成本:5 秒输出示例
路线与分辨率 | 单价 | 5 秒输出费用 | 证据 |
AutoDL Art 托管工作流,480p 高峰 | ¥0.04/秒 | ¥0.20 | 两条真实样片分别实扣 ¥0.20 |
AutoDL Art 托管工作流,480p 空闲 | ¥0.03/秒 | ¥0.15 | 页面报价,未在空闲时段实测 |
AutoDL Art 托管工作流,768p 高峰/空闲 | ¥0.05/¥0.04 每秒 | ¥0.25/¥0.20 | 页面报价,未实测 |
MiniMax 官方 H3-Lite API,480P | $0.025/秒 | 约 ¥0.84 | 官方目录价换算,未实测 |
MiniMax 官方 H3 API,768P | $0.08/秒 | 约 ¥2.69 | 官方目录价换算,未实测 |
2026-10-26 汇率按 100 美元 = ¥671.62换算;MiniMax 官方定价的输入视频和超额图片可能另收费。不同型号与工作流的画质、排队和处理过程没有做同条件对照。所有模型、分辨率、素材附加费和 AutoDL Art 18 个工作流的报价见价格矩阵。价格会变动,付费前以平台当前页面和账户预算为准。
文档与边界
使用说明书:账户准备、令牌、安装、完整调用流程、成本和故障处理。
价格矩阵:官方 H3 全型号与 AutoDL Art 18 个工作流。
验收记录:两条真实出片与费用证据。
私有 ComfyUI 备用路线:需自行租 GPU,当前只通过本地模拟测试,未完成真实出片验收。
公开仓库的 MIT 许可只覆盖本仓库代码。MiniMax H3 权重另受 MiniMax H3 Community License 约束;自行部署模型或向第三方提供服务前,应核对现行许可与使用地区。不要公开 API 令牌、签名下载地址或未经同意的素材。
Available Tools
12 toolsh3_art_download_videoB
Download a completed managed H3 MP4 into H3_OUTPUT_DIR.
Inputs: task_id (required).
h3_art_generateC
Submit one paid managed H3 text-to-video task; requires AUTODL_ART_TOKEN and H3_ART_ENABLE_PAID=1.
Inputs: prompt (required), seed, duration, resolution.
h3_art_generate_workflowB
Submit one allowlisted paid H3 workflow. Media must be public HTTPS URLs; no local file upload.
Inputs: workflow_id (required), mode, seed, audios, +6 more.
h3_art_job_statusC
Check any managed H3 task status. processing_seconds is task time, not clip length.
Inputs: task_id (required).
h3_art_list_workflowsA
List 18 reviewed AutoDL Art H3 workflows, categories, media requirements and resolutions.
Inputs: no named parameters. Returns: result
h3_art_quoteB
Estimate AutoDL Art managed H3 price using rates observed 2026-10-26; no charge.
Inputs: duration, resolution.
h3_art_quote_workflowB
Estimate one selected H3 workflow using observed rates; no charge or live price lookup.
Inputs: workflow_id (required), duration, resolution.
h3_cancel_jobB
Remove a queued H3 job or request interruption of that exact running job.
Inputs: job_id (required).
h3_check_backendA
Read the private ComfyUI system status before submitting paid GPU work.
Inputs: no named parameters.
h3_download_videoB
Download a finished H3 video through the private tunnel into H3_OUTPUT_DIR.
Inputs: job_id (required).
h3_generateB
Queue a MiniMax H3 text-to-video clip with generated stereo audio; returns job_id immediately.
Inputs: prompt (required), seed, steps, width, +2 more.
h3_job_statusB
Check whether a queued H3 video has finished and list its ComfyUI output files.
Inputs: job_id (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 12 tools
The set contains two parallel stacks (h3_art_* for managed AutoDL Art and h3_* for the private ComfyUI tunnel), producing near-duplicate tools like h3_art_job_status/h3_job_status and h3_art_download_video/h3_download_video. Descriptions clarify the managed vs. private distinction, but an agent must read carefully to avoid picking the wrong backend.
All names are snake_case with an h3_ prefix, but the naming mixes noun-first forms (h3_art_job_status, h3_job_status) with verb-first forms (h3_check_backend, h3_generate, h3_cancel_job). The two prefix families are predictable internally, yet the overall convention is not uniform.
12 tools is well within a reasonable range for a video-generation server covering quote, generate, status, download, and cancel flows. There is some duplication from maintaining both managed and private backends, but each tool is plausibly needed.
Core lifecycle is covered: pricing quotes, workflow discovery, generation, job status, download, and cancellation, plus backend health checking. Minor gaps exist, such as no explicit cancellation for managed AutoDL Art jobs and no job-listing operation.
Maintenance
Related MCP Connectors
MCP server for Hailuo (MiniMax) AI video generation
MCP server for MiniMax H3 multimodal video generation
Remote MCP for RunComfy: ComfyUI deployments, hosted models, LoRA training. 31 tools.
AI video generation MCP server for Kling, MiniMax, Luma & Hunyuan with prompt blueprints.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceLocal-first MCP server for controlling a loopback ComfyUI instance from same-machine or authenticated LAN clients.MIT
- AlicenseAqualityBmaintenanceMCP server for MiniMax H3 multimodal video generation. Enables text-to-video, image-to-video, and audio-guided video creation with async task retrieval and batch processing.10338 PyPI2MIT
- AlicenseAqualityDmaintenanceEnables MCP clients to generate images, speech, music, video, and 3D meshes via a remote mlx-serve instance on Apple Silicon.10Apache 2.0
- AlicenseNot gradedqualityCmaintenanceEnables MCP clients to connect to and interact with ComfyUI for workflow execution and image generation.Apache 2.0