Introduction

DeepSeek Harness (DSH)’s plugin architecture requires developers to decompose various capabilities into independent plugins. In local development scenarios, unifying Obsidian-style knowledge spaces, ComfyUI workflow management, and local Comfy MCP (Model Context Protocol) interactions within a single interface is a common engineering requirement. dsh-visual-workbench is a local-first visual workbench plugin for DSH. It integrates knowledge-space facets, the Comfy MCP workflow shelf, and local ComfyUI operations into DSH’s conversation view, providing a unified workbench.

What Is This

The plugin is maintained by Destined-at-Dawn and uses the MIT license. It builds a local workflow environment across three layers:

  1. Static manifest and gates: includes the static dsh-std v0.15 manifest and offline gate configuration.
  2. DSH Host: exports Cordis apply(ctx), registers the two model tools dsh_visual_workbench_status and dsh_visual_workbench_run, and provides same-origin loopback state, preview, and run routes.
  3. DSH Browser Client: injects a workbench tab through the conversation.view slot, displaying Obsidian-style facets, a bounded wikilink graph, a JSON workflow shelf, Comfy MCP status, and a visual configuration entry.

The plugin only uses the official TypeScript MCP client to connect to comfy-mcp, and does not copy the Comfy MCP source code.

Core Features

The plugin primarily provides the following capabilities:

  • Local Comfy MCP visual configuration and testing: configure local connections and port overrides directly in the UI, with support for connection testing.
  • ComfyUI lifecycle management: provides a “Start ComfyUI” button, and supports workflow preflight checks, reading the node directory, and reading model files.
  • Security safeguards: built-in safety guards include loopback address checks, token boundaries, path boundaries, and spending protection (spending is not confirmed by default).
  • Knowledge-space integration: reads the facets and bounded wikilink graph from the configured workspace, as well as up to 40 workflow JSON files.

Installation and Activation

Using this plugin requires a suitable Node.js version (Node >= 22).

The steps to install and build the plugin are as follows:

pnpm install
pnpm check
pnpm build

After the build completes, dsh.bundle.patch in package.json will cause the DSH loader to insert cordis.patch.yml, and dsh.client will provide the browser-side lib/client.js. Different DSH distribution profiles may have different dependency versions, so peer dependencies should be adjusted according to the host’s actual package versions.

Typical Usage

Using this workbench typically requires preparing the environment first, then configuring it.

  1. Prepare Comfy MCP: follow the official instructions for Comfy-Org/comfy-mcp to install comfy-mcp and comfy-cli, start local ComfyUI, and confirm that server_info is available.
  2. Configure the DSH profile: set the required environment variables. For example:
DSH_WORKBENCH_ROOT=/absolute/path/to/your/workspace
DSH_COMFY_MCP_COMMAND=comfy-mcp
DSH_COMFY_MCP_ARGS=[]

Note: DSH_COMFY_MCP_ARGS must be a JSON array containing at most 8 strings.

  1. Visual configuration and connection: in DSH’s Visual Workbench, go to Local Comfy MCP Setup, enter the absolute path and optional local port override. Click “Apply configuration and connect”, then click “Test connection”.
  2. Execute local operations: after a successful connection, click “Start ComfyUI” in Local Operations, preflight the workflow, or read models.

Use Cases and Notes

This plugin is suitable for developers who need to directly invoke local ComfyUI workflows during DSH Agent conversations and want to uniformly manage the knowledge graph and configuration.

Notes:

  • Configuration freshness: the configuration in the UI is temporary (in-process); it does not echo back COMFY_API_KEY, nor does it write long-lived credential files. For long-term persistence, configure it manually according to the official documentation.
  • Security boundaries: the default value passed is confirm_spend=false. Comfy MCP will only allow potentially paid partner nodes when the user explicitly provides true in the UI or in model tool parameters. The plugin does not wrap the Boolean value entered by the Agent.
  • Validation scope: pnpm check can only demonstrate that JSON shape, source hygiene, TypeScript types, and deterministic tests pass; it cannot prove DSH profile acceptance, ComfyUI executability, or that GPU results are genuinely valid.
  • License: Comfy MCP is a separate project, and its license boundaries are defined by its repository; the plugin itself is licensed under MIT.

Brief Conclusion

dsh-visual-workbench lowers the configuration barrier for local Comfy MCP in DSH environments through visualization, integrating local tool invocation with knowledge management. You can examine the source code or directory structure to understand its implementation details.