dsh-tool-comfyui: Let DeepSeek Harness Call Your ComfyUI Fleet to Generate Images in Chat

What Does It Do?

ComfyUI Fleet Tool is a DSH plugin for DeepSeek Harness that brings local or LAN ComfyUI image generation into the chat. It solves the waiting problem: ComfyUI uses a two-stage protocol where you submit a prompt, get an id, then poll history until the image appears, which plain fetching cannot do, so the plugin owns the polling loop. Once enabled, you just describe the image you want and the agent calls the right tool to probe workers, generate, or manage queues. Maintained by yuehancn under the MIT license, it talks only to the addresses you configure and needs no API key.

Core Features

• Fleet status probing: one call reports which workers are online, their ComfyUI version, GPU model and VRAM, queue depth, plus the available workflow templates.
• Template-based generation: submits a prompt against a named workflow template, waits for the GPU, downloads the finished image locally, and returns the file path.
• Queue management: inspects running and pending jobs on a given worker, or interrupts and clears its queue in one step.
• Multi-worker configuration: add a new machine by appending an entry to the worker list, with per-worker timeouts.
• Local-only privacy: prompts go to your own ComfyUI, images land in your output directory, with no telemetry, no credential handling, and no lifecycle scripts at install time.

How to Use This Plugin?

After enabling it in DSH, just ask in plain language: "make me an e-commerce shot of a white coffee mug" triggers generation with a matching English prompt, "is the asset machine alive?" triggers a status probe, and "clear the queue on that box" triggers queue cleanup. Before generating, register your fleet in the configuration: give each worker an id, URL, and label, then set the workflow template directory, output directory, default template, batch limit, and timeout. Adding a machine means appending one entry. On asset machines you can start ComfyUI by double-clicking the provided launcher, which listens on localhost by default or on the LAN address when you need cross-machine calls; it also stops mining processes first so the GPU is not saturated. Note that clearing a queue is destructive and should only be used when explicitly requested.

How to Troubleshoot This Plugin?

If generation is extremely slow, check whether a mining process is saturating the GPU on the asset machine; the official launcher terminates such processes before starting. If the tools never appear in chat, verify the plugin is listed in the profile's bundles array, since appearing only in dependencies does not count as enabled; fix it per the official instructions and restart that profile. If behavior looks off, check the plugin spec against your own DSH version, as this plugin was built and tested against a preview release.