Modlens
liustack
A vision bridge for text-only harnesses — images go out to a vision engine you configure, structured evidence comes back.
model
Model adapters, routing, fallbacks, cost tracking and reasoning controls.
This directory lists 3 DeepSeek Harness plugins for Models & Reasoning. Their verification levels run L1 0, L2 1, L3 2, L4 0, L5 0: no listing here has been installed or run in a sandbox, so none claims L4 or L5. 0 have passed an automated security scan, and the other 3 are marked unverified rather than safe. Declared compatibility targets are ^0.1.6-alpha.2 || ^0.1.6-alpha.1 || ^0.1.5-rc.2 || ^0.1.5-rc.1 || ^0.1.5-alpha.1 || ^0.1.3-alpha.2 || ^0.1.3-alpha.1 || ^0.1.2-rc.1 1, 0.1.0-rc.8 through 0.1.5-rc.2 1, unstated 1. Together these 3 listings carry 5,369 GitHub stars, led by Modlens at 3,985. Licences are mainly MIT 3, and the code is written mostly in TypeScript, JavaScript. Every entry records its source repository, licence, declared harness target and a copy-ready install command; the most recently checked is Command Code Provider (2026-09-18). A verification level measures how far the install path was exercised — it is not a safety guarantee, so read the source before you install.
Model plugins change which model answers and how: adapters for third party endpoints, routers that pick a model per task, fallback chains when a provider fails, token and cost accounting, and reasoning effort controls.
Read the routing rules before you rely on them. An adapter that silently falls back to a cheaper model can change output quality without telling you, and cost plugins are only useful if they attribute usage to the provider that actually served the request.
Write down the policy you want before shopping: which models are allowed, what happens on provider failure, and whether cost is hard-capped. Then find the plugin that implements that policy visibly. Silent fallback is a bug for anything that affects quality gates.
Credential handling is non-negotiable. Keys belong in profile config or environment, not in prompts the transcript keeps. For cost widgets and balance trackers, check whether the number is live API truth or a local bookkeeping guess.
Mature adapters document base URLs, auth, and failure modes. Thin ones are a single endpoint field with no timeout story. Routing suites that also inject skills should be reviewed for both surfaces: a model router that also rewrites agent behaviour is two plugins wearing one name.
Assuming “OpenAI-compatible” means identical behaviour for every parameter. Letting a free public endpoint become production load-bearing. Installing a cost tracker that needs the same broad key as the model it is watching.
If the model is simply wrong for the task, no router fixes that. If the real problem is context quality, start with context inspection rather than another endpoint. If the pain is “the agent never stops,” that is orchestration discipline, not model selection.
Routing also multiplies trust boundaries. Every endpoint you allow is another place a prompt and its tool output can land. Prefer fewer, named providers with explicit failure behaviour over a dozen free tiers chained “just in case.”
liustack
A vision bridge for text-only harnesses — images go out to a vision engine you configure, structured evidence comes back.
ysr666
Eyes for text-only agents — an opt-in vision turn with pixel-level tools, and an anonymous keyless fallback that is not private.
Mars-Sea
Unofficial provider plugin for the paid Command Code relay — live model catalog, plan-aware picker, OAuth sign-in, account rotation.
AllDSH uses Google Analytics to see which pages are read, which plugins people look for and whether an install command or a badge snippet is copied. It stays off until you allow it, and advertising storage is never used. Your choice is kept in your browser only.