Minion's pitch is that the floor matters: about 625 tokens for a bare hey, against the 20K–50K other harnesses spend before you have said anything — and that floor is paid on every single turn. For someone running a self-hosted model, that is the difference between an agent that fits the context window they actually have and one that does not.
What makes it hold up in practice is the degradation path. When a server cannot do native tool-calling, minion parses [minion_tool_call] blocks out of the model's text, escapes those delimiters before tool results re-enter context, and renders a separate reasoning_content stream as a dim thinking block. Demanding a perfect server would exclude exactly the open models minion exists to serve.
Proposal: OrcaRouter as an optional source
I'm an engineer on the OrcaRouter team, and I would like to propose OrcaRouter as an optional source for minion. It would not replace or change any existing provider — the local llama.cpp / vLLM / SGLang default, zai, together, and openrouter all stay as they are, and OrcaRouter would be opt-in behind its own key.
OrcaRouter exposes an OpenAI-compatible API and uses standard API-key authentication, so it lines up with the abstraction minion already has. Source takes a base URL, an API key, and a model id, and _discover_sources() already auto-registers key-gated aggregator sources (together, openrouter) at the end of SOURCE_ORDER so they never displace the startup source. An OrcaRouter entry would follow that same shape. To be clear, this is a proposal, not code that exists or has been tested.
Where it would help minion users
- One endpoint, many models. Chat, reasoning, image, and video models behind a single OpenAI-compatible base URL, so
/source orcarouter <model> gives a multi-model host the per-switch model override together and openrouter already support.
- Automatic model routing and provider failover. This matters most for remote sources: one upstream going down should not end a turn mid-task.
- Usage tracking and budgets. Complements the per-call token accounting minion already persists in the session JSON and can push to
MINION_METRICS_URL.
- Team access controls, for the self-hosted crowd sharing one gateway.
Ecosystem and disclosure
OrcaRouter is already appearing in open-source projects including goose, promptfoo, Dify, and RAGFlow, so there is a working integration pattern to follow.
Full transparency: OrcaRouter runs an optional open-source partner program in which approved projects can receive a 5% revenue share on OrcaRouter usage attributed to their integration. Participation is not a prerequisite for the integration — I am glad to follow whatever disclosure or governance rules you prefer.
I would welcome your thoughts on whether this fits minion's direction. If you are open to it, I would be happy to put together the implementation PR. More on the program at https://www.orcarouter.ai/built-with.
Minion's pitch is that the floor matters: about 625 tokens for a bare
hey, against the 20K–50K other harnesses spend before you have said anything — and that floor is paid on every single turn. For someone running a self-hosted model, that is the difference between an agent that fits the context window they actually have and one that does not.What makes it hold up in practice is the degradation path. When a server cannot do native tool-calling, minion parses
[minion_tool_call]blocks out of the model's text, escapes those delimiters before tool results re-enter context, and renders a separatereasoning_contentstream as a dim thinking block. Demanding a perfect server would exclude exactly the open models minion exists to serve.Proposal: OrcaRouter as an optional source
I'm an engineer on the OrcaRouter team, and I would like to propose OrcaRouter as an optional source for minion. It would not replace or change any existing provider — the local llama.cpp / vLLM / SGLang default,
zai,together, andopenrouterall stay as they are, and OrcaRouter would be opt-in behind its own key.OrcaRouter exposes an OpenAI-compatible API and uses standard API-key authentication, so it lines up with the abstraction minion already has.
Sourcetakes a base URL, an API key, and a model id, and_discover_sources()already auto-registers key-gated aggregator sources (together,openrouter) at the end ofSOURCE_ORDERso they never displace the startup source. An OrcaRouter entry would follow that same shape. To be clear, this is a proposal, not code that exists or has been tested.Where it would help minion users
/source orcarouter <model>gives a multi-model host the per-switch model overridetogetherandopenrouteralready support.MINION_METRICS_URL.Ecosystem and disclosure
OrcaRouter is already appearing in open-source projects including goose, promptfoo, Dify, and RAGFlow, so there is a working integration pattern to follow.
Full transparency: OrcaRouter runs an optional open-source partner program in which approved projects can receive a 5% revenue share on OrcaRouter usage attributed to their integration. Participation is not a prerequisite for the integration — I am glad to follow whatever disclosure or governance rules you prefer.
I would welcome your thoughts on whether this fits minion's direction. If you are open to it, I would be happy to put together the implementation PR. More on the program at https://www.orcarouter.ai/built-with.