DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Adds voice synthesis to the DeepSeek Harness Web UI, letting the agent speak replies in a cloned voice via Alibaba Cloud cosyvoice models, with a player bar, settings page, and boot sound.
Monitor upstream model catalogs across providers, select new models to add to DSH with one click, with multi-protocol auto-routing and pre-write backups.
Manages the full lifecycle of local llama.cpp GGUF models with one-click load, switch, and unload behind a stable OpenAI-compatible gateway for DSH.
Adds schema-driven inline setting rows (thinking mode, context window, timeouts, etc.) inside each provider card on the DeepSeek Harness Models page, with no hardcoded lists.
Groups the model menu by provider with slide-in panel or inline fold navigation, double-click any source to switch modes, pure local DOM enhancement with zero network requests.
Pulls live model catalogs from endpoints and auto-fills context windows, token limits, and reasoning efforts for DSH routes without patching source code.
A collection of DeepSeek Harness web plugins offering AI-suggested next prompts, logged-in Chrome browser automation, and one-click code-block execution in an integrated terminal.
One-command generation of a company-grade project handover handbook from a real repository, with evidence-backed claims, deterministic inventory scanning, stack-aware playbooks, and a strict automated quality gate.
Adds voice input to DeepSeek Harness: tap the mic, speak, and the transcript lands directly in the input box. Supports browser speech recognition and OpenAI-compatible APIs, with one-click local Whisper setup.
Trims history images to a budget before sending, keeps the latest N and replaces the rest with recoverable markers; on upstream 400 it parses the image limit and retries with fewer images.
Guides the AI to default to Mermaid diagrams in development conversations, with prompt guidance, syntax validation, output safety gating, and versioned diagram persistence via Mermaid Vault.
Converts image formats on the fly before requests reach local vision models like LM Studio, fixing the 400 base64 encoding error in DeepSeek Harness.