DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A multi-provider LLM adapter for DeepSeek Harness driven by the models.dev catalog, featuring native SSE streaming, dynamic routing, and zero pi-ai dependency.
MiMo v2.5 LLM adapter for DeepSeek Harness with thinking chain, tool calling, multi-turn conversation, and SSE streaming support.
Scenario-aware model routing for dsh: auto-detects conversation intent, selects the best-fit model, and fails over along a candidate chain with GPT as last resort.
Enables agents to manage a SiYuan Note knowledge base via the kernel HTTP API, offering 30+ tools for note CRUD, full-text search, SQL queries, block-level operations, and daily notes.
A DSH plugin that lists all OpenRouter free models, ranks them by capability, and switches session or global models in one click, with automatic reasoning-effort validation.
A DeepSeek Harness plugin that observes model-inference routes through the Sleev gateway, records usage and effective token volume, and exposes observer settings in the Web UI.
A local LLM inference control panel: unified service registry, VRAM validation, parameter profiles, and one-click benchmarking for llama.cpp, SGLang, and vLLM.
Adds editable speech-to-text to the DeepSeek Harness input box, powered by local models (SenseVoice, Paraformer, Nemotron) or Codex Desktop, with results written to draft for review before sending.
Detect and identify bird species in images via URL, Base64, or file path within DeepSeek Harness.
A fully-featured MCP calculator server covering 13 math modules including arithmetic, trigonometry, statistics, number theory, complex numbers, matrices, finance, unit conversion, and geometry.
A comprehensive image toolkit for DeepSeek Harness covering OCR, ID document recognition, multi-category object detection, general recognition, pose estimation, and instance segmentation, with URL, base64, and file path input support.
Detects human heads and bodies in an input image, returning bounding boxes, confidence scores, and labels via file URL, BASE64, or local file path.