nodecode-llmapi

LLM API: Claude, GPT, Gemini, Kimi and more on one key, each model sent on the wire it works best on. A Nodecode rewrite of llmapi-ai/hermes-provider-llmapi by llmapi-ai.

LLM API is one gateway with one key, served at https://api.llmapi.ai/v1. With this cell, /connect offers LLM API, /models lists the chat models the gateway currently serves, and a turn on one of them goes out the way the gateway wants it:

a claude-* model rides the Anthropic Messages API (/v1/messages), where the lane adds its prompt-cache breakpoints, so a tool loop reads its prompt from the cache on the second round
a GPT-5-or-later model (gpt-5*, gpt-6*, ... — not gpt-4.x, not gpt-oss) rides the Responses API (/v1/responses), where reasoning items and tool calls round-trip natively
everything else rides Chat Completions (/v1/chat/completions)

The picker, its context windows, image input, reasoning rungs and list prices come from the gateway's own public catalog, GET /v1/models, which needs no key. It is read once at start and hourly after that, and kept in memory: a gateway that does not answer leaves the last list standing, and a box that has never read one shows the nine featured models the original named.

/llmapi usage reads GET /v1/credits and prints the account's balance, its billing mode and the monthly cap, if one applies.

Get it

Tell the agent "add the llmapi cell", or run nodecode add nc://llmapi.

Set it up

Make a key at app.llmapi.ai/dashboard/api-keys, then pick LLM API in /models and paste the key when the picker asks, or set LLMAPI_API_KEY in the environment. A key saved with /connect wins over the variable, and no other provider's key is ever sent to this gateway: with neither, a round goes out keyless and the gateway says why.

The address is optional, for a regional gateway or a relay:

{
  "llmapi": {
    "base_url": "https://api.llmapi.ai/v1"
  }
}

The catalog, /llmapi usage and every model request follow it.

Use it

Ask for a model — /models lists what the gateway serves right now — then talk as usual: the wire is picked from the model's name, so mixing a Claude model and a GPT-5 model in one session costs nothing but a different lane.

/llmapi            the address, whether a key is set, and what the catalog holds
/llmapi usage      the credit balance, billing mode and monthly cap
/llmapi models     how many models the picker offers, and when they were read
/llmapi refresh    read the catalog again now

(help :llmapi) answers the same from the model's side.

A model with a window under 64K is not in the picker but still runs when asked for by name: the floor is the one the original set, not a limit of the gateway.

Test

From this repository's root:

.github/test-cell.sh providers/nodecode-llmapi