dsh-mcp-skill-control

Managing MCP servers in DeepSeek Harness usually means hand-editing YAML: adding a server, toggling a connection, or tracking down why a row has zero registered tools all happen outside the UI, slowing iteration down and inviting syntax errors. dsh-mcp-skill-control brings that control panel into the dsh web session header. The MCP Servers tab lists every server with its transport, endpoint, live connection status, and tool inventory, then lets you disable, enable, restart, add, or remove entries with simple buttons. A JSON import flow accepts configs from Claude Desktop, Cursor, or OpenCode, automatically slugifying names, rejecting unsupported SSE endpoints, and skipping entries marked disabled in the source document.

The Skills tab aggregates local skills from ~/.dsh/skills and ~/.agents/skills into a single view. Toggling a skill on or off edits only the disable-model-invocation field in frontmatter while preserving every byte of the Markdown body. For streamable-http servers that appear active yet register zero tools, the plugin fires a one-shot MCP initialize probe (rate-limited to 30 s) and surfaces a human-readable diagnostic instead of leaving the row stuck in a perpetual connecting state.

The plugin ships as a single package with both a Host half and a Browser half, requiring zero changes to DSH source code. All configuration writes go through a YAML Document AST editor so user comments, key order, and inline arrays survive untouched; changes are applied in-process via HMR transactions and persist across restarts. It is ideal for DSH users juggling multiple MCP servers, frequently swapping skill combinations, or migrating a Claude Desktop / Cursor setup into the DSH workflow.