dsh-plugin-llmtr

Juggling multiple AI provider keys, model catalogs, and data-residency rules is a growing burden as teams widen their model usage. dsh-plugin-llmtr plugs the LLMTR gateway, an OpenAI-compatible AI aggregator, into DeepSeek Harness so a single API key unlocks 170+ chat models from OpenAI, Anthropic, Google, Qwen, Mistral, xAI and others, plus models LLMTR hosts on its own infrastructure in Turkey for deployments that need prompts to stay in-country. The plugin registers those models as the llmtr provider route on ctx.llm, surfacing them in the model picker where they can be selected per session, per agent, and per subagent with no manual catalog maintenance.

What makes the integration practical is that the catalog is live: the model list is pulled from the gateway on demand, so new models LLMTR adds appear automatically without a plugin upgrade. Deployments can narrow the catalog scope to only Turkey-hosted models, guaranteeing that no session starts on a model whose prompts leave the country, or pin an exact list of model ids to fully replace the live listing. Reasoning output from thinking models is read transparently with no configuration, while reasoning-effort selection is opt-in and scoped to the models that actually accept it; undeclared efforts are refused by name rather than silently dropped. Privacy is straightforward: the plugin sends only a standard User-Agent, an x-llmtr-client header, and the API key, with no session id, prompt text, file path, or user identifier added to any header.

The plugin fits teams already running DeepSeek Harness who want a broader multi-provider model catalog behind one key, and is especially relevant to organizations with Turkey data-residency obligations or any mandate to consolidate provider access through a single gateway. It is scoped strictly to streaming text chat completions, leaving embeddings, image, and other endpoints out of reach, which keeps the integration surface small and predictable.