nodecode-liquid
Liquid Inference: an auction exchange for LLM inference. A Nodecode rewrite of architect-xyz/hermes-liquid-inference by architect-xyz.
Each request goes to the eligible offer with the lowest estimated charge,
and carries a cap on what that request may be charged. This cell puts the
exchange in Nodecode as the provider liquid:
/connectoffers Liquid Inference and/modelslists its model,liquid.auto- every request carries
x-liquid-cap-usd, the most that request may be charged: 1.00 USD unless you say otherwise - every request carries
X-Title, naming the client it comes from - a reasoning level you pick rides as a top-level
reasoning_effort; a turn with none sends none
The exchange speaks the OpenAI chat wire, so a turn on liquid.auto goes to
https://router.inference.ai.exchange/v1/chat/completions like any
catalog provider's.
Get it
Tell the agent "add the liquid cell", or run nodecode add nc://liquid.
Set it up
Make a buyer key with the spend scope at
https://inference.ai.exchange. Save it with /connect liquid, or set
LIQUID_API_KEY in the environment. A saved key wins over the variable.
Then the section, a top-level key of ~/.nodecode/config.jsonc:
{
"liquid": {
// where the exchange is served; a relay's address goes here
"base_url": "https://router.inference.ai.exchange/v1",
// the most one request may be charged, in USD
"cap_usd": "1.00"
}
}
The section is on unless it says "enabled": false. With it off, a turn still
reaches the exchange on a key, but the request carries neither the cap nor the
client name.
Use it
nodecode config set model.provider liquid nodecode config set model.default liquid.auto
/liquid says how the exchange stands: its base, its cap, its model, and
whether a key is there. A cap that is not a positive decimal ("abc",
"0.00") refuses at start, because a malformed cap is the uncapped bill the
header exists to prevent.
Test
From the cells repository's root:
.github/test-cell.sh providers/nodecode-liquid
Offline: every wire is stubbed.
