OpenAI-Compatible Provider

This plugin connects PeckBoard to any endpoint speaking the OpenAI /v1/chat/completions protocol — OpenRouter, LM Studio, vLLM, llama.cpp, and anything else with a compatible server. It is the no-code way to add a model source PeckBoard does not ship a provider for; the Providers page covers the first-party providers it sits alongside.

Point its settings at the API root (base_url, e.g. https://openrouter.ai/api/v1 or http://127.0.0.1:1234/v1), list the model ids to expose, give the provider a display name for the model picker, and supply an api_key if the endpoint needs one — local servers accept an empty key. The models then appear in every model picker as openai-compat:<model> under your chosen display name, and sessions on them behave like any other: streamed events, usage, and cost all flow through the normal pipeline.

Hooks and behaviour under failure

Hooks: provider.register, provider.models, provider.send. As a provider plugin it owns whole agent turns: a turn that times out, traps, or ends without a terminal event is surfaced as a crash by core, so a session never wedges. Interrupts are cooperative — the plugin checks a stop flag between chunks. Changing its settings can re-trigger the approval prompt, since approval is bound to the exact hook and permission set.