DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A local multi-channel AI gateway that unifies Google and OpenAI subscription accounts into a standard OpenAI-compatible API, with multi-account rotation, real-time quota dashboards, and adaptive egress proxy.
Auto-fills missing capability metadata (context window, output cap, reasoning levels, vision) for custom-gateway models by name, without modifying your config.
Integrates Cline Pass into dsh with every request pinned to the DeepSeek upstream channel (no fallback); API key managed in Settings, 5-hour/weekly/monthly usage bars on the model card, no local port opened.
An autonomous screen-capture plugin for DeepSeek Harness that lets the agent see the screen in the same tool call, with multi-frame, region, multi-display support and macOS permission guidance.
Thin dsh plugin for workbuddy2api: manages gateway lifecycle, auto-syncs model catalog, and installs with one command, no Go or Python required.
Podman-backed execution layer for DeepSeek Harness that routes all shell commands and file access through disposable, per-project containers.
An end-to-end analysis toolkit for empirical asset pricing and machine learning, covering literature review, data processing, model training, paper writing, and presentation creation.
Preset Bailian (DashScope) inference route with 36 text models plus an auto-adapter that fills in missing dialect, reasoning levels, and capacity declarations without hand-written config.
Adds the MotoMoto relay route to the model selector via the llm-pi-ai adapter, proxying the OpenAI Responses API with Codex CLI identity and a live status card in plugin settings.
A DeepSeek Harness plugin for general object recognition, identifying the main subject in images across 50,000+ categories via URL, Base64, or file path.
A DSH plugin that humanizes Chinese writing by removing AI-specific words, breaking template phrases, cutting vague closings, and auto-learning avoidance patterns.
Gives text-only models a pluggable vision sidecar by rewriting image blocks into recognition prompts handled by a configurable VLM, while keeping original images in the UI and history.