DeepSeek Harness Plugins · Model Inference
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
Sign in to Claude, Codex, Grok, Qwen Code, and other AI subscriptions via OAuth and use them as LLM providers in DeepSeek Harness, no API keys required.
Registers the self-hosted Lynkr multi-provider LLM tier-routing gateway as a custom OpenAI-compatible provider in DeepSeek Harness, automatically routing each request to the cheapest capable model tier.
Integrates locally logged-in CodeBuddy/WorkBuddy subscriptions as a native provider in DeepSeek Harness, supporting SSE streaming and function calling without separate authentication.
dsh-commit-ai is a model-inference plugin for DeepSeek Harness that leverages AI to assist with commit-related tasks.
A Settings-based UI to configure per-model input modalities (text/image) and reasoning effort levels across all providers and overrides, with no manual settings.yaml editing required.
Enables text-only models to recognize images via free Zhipu GLM vision models, with credential-aware routing and panel-based API key setup.
A DSH reasoning-effort slider plugin with 7 levels, auto configuration, whale-fall theme, chibi animations, self-healing, and compatibility fixes.
Provides a unified ontology interface for DeepSeek Harness with built-in local storage and browser projection adapters for versioned registration of objects, links, functions, and actions.
Reuse a Google Antigravity subscription in DeepSeek Harness with multi-account switching, serving Gemini, Claude, and GPT-OSS models through a local OpenAI-compatible proxy.
A thin LLM adapter that registers a local llama.cpp server as a standard Harness provider via its OpenAI-compatible /v1/chat/completions endpoint, supporting streaming text and tool forwarding.
Integrates 12 strongly-typed Talebook tools into DSH for book management, audios, remote sources, and admin operations with built-in risk gates and verification.
Automatically adjusts reasoning effort per task using keyword heuristics, explicit tags, and message length, with peak-hour capping and on-error escalation to avoid a single global setting.