AtlasCloudAI
dsh-media-gen
Plan and execute Atlas Cloud image, video, audio, and 3D workflows in DeepSeek Harness.
- Stars
- 0
- Language
- JavaScript
- Created
- Aug 14, 2026
- Updated
- Aug 17, 2026
Introduction
dsh-media-gen
Turn one media brief into a model choice, production-ready prompt, and—when you opt in to MCP—an executable Atlas Cloud workflow.
dsh-media-gen is an independent DeepSeek Harness profile bundle maintained by AtlasCloudAI. It covers image, video, audio, and 3D workflows. The three Skills become discoverable after installation; the optional Atlas Cloud MCP bridge is included but disabled by default.
Quick start
Install the bundle into the DSH profile where you want to use it:
dsh plugin --profile web add 'github:AtlasCloudAI/dsh-media-gen#v0.2.0'
dsh --profile web --dump-config
Use --profile headless instead for a headless profile. Pin a release tag or full commit for reproducible installations.
Try it
Start that DSH profile and paste this no-submit demo:
Plan an 8-second coffee product video. Compare Seedance with one available alternative, recommend a model, and produce a three-shot storyboard, final prompt, and exact parameters. Do not submit generation; wait for my confirmation.
Depending on the request, DSH can use the bundled Skills to compare model approaches, turn the brief into a coherent storyboard, and prepare the selected model's input. The demo does not ask DSH to submit a generation request.
What each component does
| Component | Problem it solves | Typical result |
|---|---|---|
atlas-cloud Skill | Which Atlas Cloud model, API, and schema should I use for image, video, audio, 3D, ASR, or LLM work? | A model ID, validated parameters, and a REST, CLI, or MCP execution path. |
seedance-2-5-skill Skill | How do I plan a controllable, consistent Seedance video across shots and references? | A storyboard, continuity plan, and Seedance-ready prompt and parameters. |
universal-video-prompt-skill Skill | How do I reuse one video brief across different generation models? | One model-neutral prompt specification plus model-specific compilations. |
Optional atlascloud-mcp@1.5.0 | How can DSH call supported Atlas Cloud operations without hand-wiring each request? | Tools for model and schema lookup, media upload, generation, polling, and account usage checks. |
In short: Skills teach DSH how to plan and integrate; MCP gives it callable execution tools.
Credentials and execution
Installing and discovering the Skills requires no Atlas Cloud credential and submits no Atlas Cloud API request.
For execution, obtain a key from the Atlas Cloud console and set it in the process that starts DSH:
export ATLASCLOUD_API_KEY="<your-key>"
Do not paste the key into chat or commit it to this repository.
Enable MCP execution (optional)
The bundle's atlascloud-mcp row is disabled by default because a stdio MCP server is a trusted child process that runs outside the agent sandbox.
To opt in, add this later-layer override to the target profile's $DSH_HOME/profiles/<profile>/cordis.patch.yml:
- id: atlascloud-mcp
disabled: false
Restart the profile and inspect the resolved configuration:
dsh --profile web --dump-config
DSH exposes qualified tool names such as mcp__atlascloud__atlas_list_models; the underlying MCP tool name remains atlas_list_models. Operations that submit generation or transcription can be billable, so review the exact model and parameters before approving a submission.
Compatibility
OpenAI/Codex plugins and DSH bundles use different host manifests. The Skill content can be reused, but the host integration cannot be installed unchanged: DSH requires package.json with dsh.bundle.patch and a Cordis MCP row. This repository uses the current DSH profile-bundle format, not the retired .dsh-plugin format.
See the full compatibility decision.
Provenance and verification
The Skills are synchronized from AtlasCloudAI/atlas-cloud-skills. skills/SOURCE.json records the exact source commit and local DSH adaptations.
npm test
npm pack --dry-run
The checks validate the bundle manifest, default-off MCP policy, pinned MCP executable, Skill frontmatter, public-language guard, source pin, and packaged relative resources. They do not submit generation, upload media, transcribe audio, or call a billable Atlas Cloud endpoint.