nc://commandcode

Command Code: the models of a Command Code account as a Nodecode provider, over the same /alpha/generate wire the Command Code CLI speaks. A Nodecode rewrite of NanamiMio/hermes-commandcode-oauth by NanamiMio.

With it, /connect offers Command Code, /models lists its models, and a turn on commandcode/<model> goes to https://api.commandcode.ai/alpha/generate the way the vendor's own CLI sends it: the CLI's client headers, the key the CLI already stored, and the CLI's six-digit-version wire.

Get it

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

Set it up

Sign in with the Command Code CLI once (~/.commandcode/auth.json then holds the key), or save a key from https://commandcode.ai with /connect, or set COMMANDCODE_CLI_TOKEN. A key saved with /connect wins over both; the CLI's file is only ever read, never written.

The section is optional and on unless it says "enabled": false:

{
  "commandcode": {
    // where Command Code is served; the lane appends /alpha/generate
    "base_url": "https://api.commandcode.ai",
    // the file the CLI keeps its sign-in in
    "cli_auth_file": "~/.commandcode/auth.json",
    // import the CLI's stored key when nothing else answered
    "read_cli_login": true
  }
}

Use it

Pick commandcode in /models, then a model: the CLI's snapshot lists DeepSeek V4/V4.1, Kimi K3, GLM 5.3 Flash, MiniMax M3/M3.1, Qwen3.7 Max and Qwen3.8 Omni Flash, the Muse Spark pair, and the zero-cost Laguna S 2.1. A turn carries the conversation, the tool schemas, the working directory and the OS/architecture to Command Code. Images ride to the models that take them.

/models asks the account's own catalog (GET /provider/v1/models?view=client) first and falls back to the snapshot when the account cannot be asked.

What the wire is

The request is the CLI's own body: a workspace config block, an agent mode, and params carrying the model, the messages as typed parts (text, reasoning, tool-call, tool-result, image), one system string, the tools as input_schema declarations, the output budget and stream: true. The answer is NDJSON: text-delta, reasoning-delta, a whole tool-call, finish with totalUsage (the cached part of the prompt apart from the rest). An in-stream {"type": "error"} on a 200 is a failure carrying its status, and a stream that ends without finish fails closed rather than handing a turn an empty answer.

Not ported

See NOTICE.

MIT licensed.