MCP Servers for Genmedia x DeepSeek Harness (dsh)

Use the MCP Servers for Genmedia with the DeepSeek harness (deepseek-ai/deepseek-harness, dsh), a Cordis “everything is a plugin” harness in developer preview. It supports MCP servers (stdio and streamable-http) through its first-party plugin @deepseek-ai/dsh-mcp-client. To install the server binaries, see Installation.

This is a developer preview, so expect breaking changes.

Add the servers as a Cordis - insert: operation in a patch file ($DSH_HOME/cordis.patch.yml, where $DSH_HOME defaults to ~/.dsh), or apply a patch ad hoc with dsh web --patch <file>. New rows must go under - insert:; a bare top-level row is the override-by-id form and will not register a new server.

- insert:
    - id: veo
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: veo
        transport: stdio
        command: mcp-veo-go
        args: []
        cwd: !!js process.cwd()
        toolCallTimeoutMs: 300000
        env:
          GOOGLE_CLOUD_PROJECT: YOUR_GOOGLE_CLOUD_PROJECT_ID
          GENMEDIA_BUCKET: gs://YOUR_GENMEDIA_BUCKET
    - id: nanobanana
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: nanobanana
        transport: stdio
        command: mcp-nanobanana-go
        args: []
        cwd: !!js process.cwd()
        env:
          GOOGLE_CLOUD_PROJECT: YOUR_GOOGLE_CLOUD_PROJECT_ID
          GENMEDIA_BUCKET: gs://YOUR_GENMEDIA_BUCKET

Tools appear as mcp__<serverName>__<tool>. Add the remaining servers (gemini, chirp3, lyria, avtool, omni) the same way, as additional list items under the same - insert: operation. Use toolCallTimeoutMs (milliseconds, default 60000) to raise the per-tool timeout for long jobs like Veo.

Important: The dsh stdio bridge strips ambient variables whose names look like credentials, and all DSH_* variables, before launching the child process. You MUST put GOOGLE_CLOUD_PROJECT, GENMEDIA_BUCKET, and any credential path in the entry’s config.env rather than relying on your shell environment.

dsh has native SKILL.md discovery via @deepseek-ai/dsh-skill-filesystem, which scans top-level roots including ~/.agents/skills, ~/.dsh/skills, and the project’s .agents/skills and .dsh/skills. Each skill directory must sit directly at one of those roots (for example ~/.agents/skills/genmedia-image/SKILL.md); dsh does not discover skills nested deeper (a **/SKILL.md under some other folder is not found). Place each genmedia skill at a scanned root to use it in dsh.

Authentication

The genmedia MCP servers call Vertex AI from the server process, so each needs Google Cloud Application Default Credentials (ADC) and a project ID. For dsh, set these in the entry’s config.env, not your shell:

gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT="$(gcloud config get-value project)"

For service accounts, set GOOGLE_APPLICATION_CREDENTIALS to the key file path (again, in config.env for dsh). GENMEDIA_BUCKET (a gs:// URI) is optional and sets the default GCS output destination. Long-running models like Veo need a raised per-tool timeout.