GitHub MCP Reference (optional)

GitHub's official github-mcp-server is an optional structured backend for this pack. Nothing in this package requires it: the four skills compose whatever GitHub/Git capabilities are already visible in the session and fall back to gh/git when needed. This reference explains how to make GitHub MCP tools visible to DSH when you want them.

This package does NOT auto-edit DSH profiles, does NOT configure MCP servers, and does NOT manage credentials. Wiring an MCP server is a profile configuration step you perform yourself (via the official dsh plugin tooling / profile file); this pack only consumes the tools that end up in the session's tool catalog.

DSH MCP client (current supported transport)

DeepSeek Harness ships an MCP client bridge, @deepseek-ai/dsh-mcp-client. The reviewed baseline (0.1.0-rc.6) supports two transports: stdio and streamable-http. Declare one plugin instance per MCP server in the profile's cordis.yml:

- id: mcp-github
  name: '@deepseek-ai/dsh-mcp-client'
  config:
    serverName: github
    transport: stdio
    command: docker
    args:
      - run
      - -i
      - --rm
      - -e
      - GITHUB_PERSONAL_ACCESS_TOKEN
      - ghcr.io/github/github-mcp-server
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: !!js process.env.GITHUB_PAT

Note on env vs container: the env block feeds the docker CLI process, not the container. To authenticate, the token must be forwarded into the container with -e GITHUB_PERSONAL_ACCESS_TOKEN (as in the server's own Docker examples); env supplies that variable to docker run. For non-Docker installs, set the token in the environment the server process itself runs under.

GitHub MCP server

• Official repo: github/github-mcp-server (docs: README.md, docs/server-configuration.md, docs/github-app-auth.md).
• Auth: OAuth, or a GitHub Personal Access Token via the GITHUB_PERSONAL_ACCESS_TOKEN environment variable (PAT takes precedence over OAuth).
• Credentials come from environment / secrets only — never hard-code a token into a profile file. !!js process.env.<NAME> keeps the secret in the environment. This package never reads tokens out of gh and never stores credentials (see references/safety-model.md).
• Prefer read-only / limited toolsets where suitable: --read-only exposes only read-only tools; --toolsets repos,issues,pull_requests,... (or GITHUB_TOOLSETS) limits the surface to what you need. The server docs note that enabling only needed toolsets "can help the LLM with tool choice and reduce the context size".
• Avoid loading every GitHub MCP tool unnecessarily: every registered MCP tool's schema is model-visible and costs context on every request while it is registered (DSH mcp-client docs: "Data-dependent schema cost is paid on every request while the tools are registered").

Verifying visibility

After the profile boots with the MCP entry, the GitHub tools appear in the session's tool catalog under server-qualified names mcp__github__<tool> (e.g. mcp__github__create_pull_request). Verify by checking that the running session's catalog actually lists the mcp__github__* tools (e.g. via the session tool list). dsh --profile <profile> --dump-config only proves that the MCP instance is configured; actual visibility must be confirmed from the running session's tool catalog. Only tools the model can actually observe are usable — a server that fails to connect or synchronize registers nothing.

How the core skills use them

Core skills choose MCP tools by documented semantics, exactly like any other visible capability: use the most specific visible tool whose documented semantics cover the need. Provider names and tool-name prefixes (including mcp__*) do not imply capability. MCP tools that mirror gh/git surfaces are fine to use when visible and semantically sufficient; otherwise the skills fall back to gh/git and this pack's bundled helpers.