DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A resident drawing panel for DSH powered by local ComfyUI, featuring dual prompts, parameter control, image preview and safe saving, with one-click deployment of a pinned ComfyUI workflow.
Integrates Zhipu GLM Coding Plan web search, page reading, and GitHub repo tools as native DSH providers with live settings.
DSH local vision plugin: paste screenshots in chat for AI analysis with OCR coordinates and semantic understanding, fully offline on Mac.
Declares reasoning tiers and image input for manually added OpenAI-compatible models so the Effort selector and image attachment become available in chat.
Retry any model-request failure in DSH with durable exponential backoff, tunable via live sliders for retry budget and 429 wait.
Let text-only DeepSeek-V4-Pro handle images by delegating vision to Flash-Vision-Exp and feeding captions back into the conversation.
Vision analysis plugin for DeepSeek Harness with 8 modes (OCR, chart-data, UI review, code-gen, etc.), a zero-config free endpoint, and support for any OpenAI or Anthropic vision API.
Direct bridge to DeepSeek's official vision API, registering deepseek-v4-flash-vision-exp as a native provider route with no third-party model involved.
One-click start/stop of local llama.cpp servers from the DSH settings page, connecting local LLMs as session models with dual slots and editable launch parameters.
Dual-model auto-routing for DeepSeek Harness with six-criteria decisions and an escape-learning loop that continuously improves routing quality.
Adds reasoning effort selectors to non-DS local gateway models in DeepSeek Harness, with presets for 20 common models and one-click drift repair.
Splits your session into a pro-powered planner and flash-powered executors, pairing strong reasoning with fast cheap implementation to cut token costs without sacrificing quality.