byok-custom-model

Registers a custom LLM endpoint with a user-supplied API key so it appears as a selectable chat model in Starchild, bypassing the platform proxy and calling the vendor or aggregator directly. It is a script-mode skill with no registered tools: the agent reads the file and then calls Python exports from a bash block. The skill ships with eleven pre-configured vendor presets — including Anthropic, OpenAI, xAI/Grok, Qwen/DashScope, DeepSeek, Kimi/Moonshot, MiMo, Gemini, Gemma, NEAR AI, and Venice — each with base_url, wire format, thinking parameters, and capabilities already filled in.

The onboarding flow is templates-first. The user's intent is matched against the curated vendor table before asking for any URL, and a match goes straight to add_template(vendor=...). Only when no curated vendor fits does the flow ask the user to paste the provider's official API example (without a real key), run parse_example to auto-detect base_url, upstream_model, wire (openai vs anthropic), thinking params, and vendor-specific request fields, and then call add(...). Whenever add() or add_template() returns a need_env_input field, the agent must immediately call the request_env_input tool, which pops a secure-input UI so the key lands in workspace/.env; the script cannot pop that UI itself. The exported functions include templates, list_vendor_models, add_template, parse_example, add, list_models, get, and remove, each returning a dict with ok True or ok False and an error.

Security rules are strict: API keys pasted in chat are refused, the secure popup is the only accepted channel, the popup is never re-issued automatically while waiting, and custom_models.yaml and .env are never hand-edited. After registration the model appears prefixed with custom/ and is selected with /model custom/<name>, with vendor pricing applied directly to the user's BYOK quota. A privacy-first tier favors NEAR AI for its shorter hardware-enclave trust chain and recommends TEE-protected models there; the skill also documents NEAR's reasoning protocol, which nests chat_template_kwargs under extra_body rather than using top-level reasoning parameters.