dsh-llm-mlx
Run the following command in DeepSeek Harness:
dsh plugin install robbywang25/dsh-llm-mlx
Paste the following prompt into your AI chat to install this plugin:
Run dsh plugin install robbywang25/dsh-llm-mlx inside DeepSeek Harness to install this plugin; the full source is available at https://github.com/robbywang25/dsh-llm-mlx
About this plugin
Running local LLM or VLM inference on an Apple-silicon Mac through the MLX ecosystem often means manually launching mlx_lm.server or mlx_vlm.server, juggling ports and flags, and then wiring the endpoint into whatever framework you are building on top of. dsh-llm-mlx closes that gap by contributing a local-mlx model route through the built-in OpenAI-compatible adapter of DeepSeek Harness, so the MLX weights already sitting on your disk appear directly in the DSH model list with no extra middleware in between.
The plugin offers two operating modes: reuse a server you manage independently (DSH never owns its process), or let DSH spawn and own the MLX server for the lifetime of its own process, cleaning up only the child it started at exit. For users who route Claude Desktop through CC Switch and hit an SSE deserialization issue caused by duplicate reasoning fields in some MLX-VLM releases, an optional loopback-only compatibility proxy strips the deprecated alias while passing every other field through unchanged. The bundle also fixes a macOS subprocess-provider path-rewrite bug in DSH Desktop 2.0.3, leaving Linux and Windows providers untouched.
This plugin is aimed at developers on Apple-silicon macOS who want offline, local-weight inference through DSH, at privacy-first users who refuse to send prompts to a remote API, and at teams that need to call a local model from Claude Desktop. It downloads no weights, uploads no prompts or telemetry, launches Python via an argument array rather than a shell, and binds every server endpoint to 127.0.0.1 exclusively.
Use Cases
- Offline LLM or VLM inference on an Apple-silicon Mac using local MLX weights
- Calling a local MLX model from Claude Desktop through CC Switch with SSE compatibility
- Letting DSH own the MLX server lifecycle to skip manual start-stop and port management
Best For
- Developers running local inference on Apple-silicon Macs
- Privacy-first users who refuse to send prompts to remote APIs
- Teams building local workflows on DeepSeek Harness
Related Plugins
ModLens is a vision plugin for DeepSeek Harness that gives text-only models sight by reading images pasted directly into chat, with zero-config setup and multiple vision engines.
On-demand vision for text-only DeepSeek Harness agents: built-in free keyless vision chain and 14 vision tools, routing image turns as tool calls to vision models with pixel fidelity, no Python needed, one-command install.
Give text-only models in DeepSeek Harness eyes, enabling image Q&A, long-screenshot OCR, UI restoration, and GUI visual tasks.
An armor-breaking plugin for DeepSeek: rewrites instruction priority, expands refusal-blocked phrases, persists across sessions, and shows a green active indicator.