dsh-unsloth-hands
Run the following command in DeepSeek Harness:
dsh plugin install MicroHEROX/dsh-unsloth-hands
Paste the following prompt into your AI chat to install this plugin:
Install the plugin in DeepSeek Harness by running dsh plugin install MicroHEROX/dsh-unsloth-hands; the full source code is at https://github.com/MicroHEROX/dsh-unsloth-hands
About this plugin
The flagship online model behind DeepSeek Harness is powerful, but routing repetitive, token-cheap jobs through it — batch rewrites, name translations, string munging, short summaries — is simply wasteful. Even trickier: most flagship models are text-only, so image analysis, OCR, and multi-image comparison are out of reach entirely. dsh-unsloth-hands fills that gap. It gives your online main model a pair of local hands, delegating high-frequency, low-stakes work to a model already loaded in Unsloth Desktop on your machine, while the online model stays in charge of decisions and final review.
The plugin registers two tools on the harness. unsloth_run sends a text prompt to the local model for batch rewrites, translations, extraction, and short summarization, returning the full result in a single round trip over the OpenAI-compatible chat-completions endpoint. unsloth_vision handles vision tasks with three built-in prompt templates — analyze (eight-section structured report), ocr (character-exact transcription), and compare (joint reasoning over two to four images) — and accepts images from local file paths, data or HTTP URLs, or conversation attachments. Before every call the plugin probes /v1/models so a missing server surfaces as a clear, actionable message rather than a generic network failure. Authentication uses a Bearer token from Unsloth Settings, and live config overrides let you adjust behavior without restarting the harness.
It is aimed at anyone already running DeepSeek Harness for daily development or content work who also has Unsloth Desktop loaded with a local model, even on a modest Mac or a small GPU box. The plugin is a pure HTTP client: it never starts, configures, or stops Unsloth, bundles no model files, and modifies no files in either DeepSeek Harness or Unsloth. Install it, create an API key in the Unsloth app, add three lines of YAML to your profile, and your online model gains the ability to quietly reach out to a local inference engine whenever a task is better done close to the metal.
Use Cases
- The online model delegates batch rewrites, translations, or short summaries to a local Unsloth model to save tokens
- The flagship model is text-only and cannot process images, so a local multimodal model handles OCR or image analysis
- Compare two to four screenshots locally and produce a structured difference report
Best For
- Users already building daily work around DeepSeek Harness
- Developers who run Unsloth Desktop locally and want to reuse their local model's compute
- Teams that want to cut token spend on repetitive light tasks while keeping the online model in the driver's seat
Related Plugins
A service-oriented sidebar framework for DSH, offering a full workbench with file explorer, embedded browser, real terminal, Git panel, and extensible plugin services.
A beautiful, practical Claude Code-style TUI plugin with pixel whale top bar, flowing glow title, real-time status line, streaming thought expansion, time rewind, context progress bar and TPS gauge — zero core changes.
The plugin market for DeepSeek Harness: browse/search/one-click install community plugins and themes, with updates, backup/restore, hot disable, diagnostics, and AI fix.
A persistent whale widget on the DSH web UI that shows DeepSeek balance, today's usage, and per-turn cost, with drag-and-snap and sound effects.