Introduction

DeepSeek Harness (DSH) adopts a plugin-based architecture, but in real development, managing MCP servers often comes with pain points such as fragmented configuration, opaque connection status, and difficult tracing of tool calls. Existing approaches may require restarting the process or relying on external tools. dsh-mcp-studio integrates these features into the DSH settings page and provides an MCP server workbench that can take effect without a restart.

What Is It?

dsh-mcp-studio is an MCP server workbench for DeepSeek Harness (DSH), maintained by SeaOf0. Its core goal is to answer how to mount, monitor, and manage MCP servers in real time within the DSH environment, and how to import and export configurations conveniently.

Core Features

  1. Real-time mounting and hot switching
    Each enabled server is bridged and mounted by the built-in @deepseek-ai/dsh-mcp-client. The model can directly call mcp__<name>__<tool>. Editing the configuration enables hot switching, while disabling or deleting automatically unmounts the server.

  2. Accurate connection status
    Status is aggregated directly from the DSH tool registry (ctx.tools.view()). Displayed states include: connected, starting, disconnected, failed (with error details), and disabled. The connecting state shows a pulse breathing-light effect.

  3. Tool preview overlay
    Click the “N tools” badge to open a standalone overlay showing the server name, total number of tools, and a brief description. Instant filtering is supported. The overlay automatically adapts to the viewport (top-bottom flip, height clamping, and scroll following).

  4. Tool call monitoring
    Subscribes to the session event stream and records the server, tool, duration, and success or failure of each call. Uses an in-memory circular buffer with a capacity of 200 entries, paginated at 20 per page. Supports filtering by server and one-click clearing with a two-step confirmation. Data is cleared after restart.

  5. Connection diagnostics
    Initiates an independent MCP handshake with one click. Supports stdio temporary child processes and direct streamable-http connections. Shows handshake duration, negotiated protocol version, server information, and tool count.

  6. One-click JSON import
    Compatible with multiple configuration file formats, including Claude Desktop, VS Code, and Cline/Roo (including disabled entries). Supports bare mappings, unwrapped single objects, and multi-layer wrapping. Automatically detects stdio/http and env/headers, and automatically adds suffixes to duplicate names.

  7. Template library
    Includes 9 commonly used server presets: Everything, Filesystem, Fetch, Memory, Seq Thinking, Puppeteer, GitHub, Chrome DevTools, and an HTTP example.

  8. Configuration export
    Supports copying the current full configuration or a single server’s configuration as JSON in the mcpServers format with one click.

  9. Sorting and filtering
    Supports card drag-and-drop sorting (persisted), searching by name/command/URL, four-state status filtering, and switching between compact and comfortable density.

  10. Design
    Uses --dsw-alias-* theme tokens and adapts automatically between light and dark skins. Hierarchy is expressed through font size, spacing, and status colors, with zero emojis and zero decorative icons throughout.

Installation and Enabling

Prerequisites:
* DeepSeek Harness (DSH) version >= 0.1.0-rc.6
* Node.js version >= 22.19 or >= 24

Installation command:

npx @deepseek-ai/dsh plugin --profile web add github:SeaOf0/dsh-mcp-studio

After installation, restart dsh web. The MCP workbench will appear in the settings dialog, located below the side cards. Configuration is persisted in the mcp-studio section of ~/.dsh/settings.yaml.

Typical Usage

  1. Add or import configuration: Click “Add Server” to configure manually, or click “Paste JSON” to import. The drawer includes template and formatting buttons. You can also click a preset chip to fill it in with one click.
  2. Invoke in conversation: Directly instruct the model in the conversation to use a specific server. For example, “Use the github server to check the xxx repository.” The model will automatically call mcp__github__* tools.
  3. View status: The statistics bar shows the number of servers, enabled servers, connected servers, and total tools in real time. A green breathing-light status on a card indicates that the server is connected.
  4. Diagnose and debug: Click the “N tools” badge to view details, or click “Test Connection” to run an independent handshake and view the protocol version and duration.

Applicable Scenarios and Notes

Suitable for DSH users and developers who need in-depth management of MCP server connections, tool invocation status, and configurations.
Notes:
* The plugin runs with the permissions of the current DSH process.
* Ensure the host package has been externalized. Runtime dependencies are resolved from ~/.dsh/profiles/node_modules.
* It is recommended to review the source code and license before installation.

Conclusion

By integrating mounting, monitoring, diagnostics, and import/export functionality into the settings page, dsh-mcp-studio addresses the fragmentation problem in MCP server management. Through hot switching and persistent configuration, it reduces management complexity.

Directory page | GitHub repository