nodecode-carbonvoice

Carbon Voice: read the messages that came into a voice-messaging workspace and answer them in the thread. A Nodecode rewrite of PhononX/hermes-plugin-carbonvoice by PhononX.

Carbon Voice turns what people say into text. The model gets four tools over the same API a Carbon Voice client speaks to, and a person gets one command:

What	Does
`carbonvoice_recent`	The messages that came in lately, newest last: who sent each one, the conversation it is in, the transcript Carbon Voice made of it, and the thread it answers.
`carbonvoice_message`	One message by id, so a voice memo whose transcript has not landed yet can be read once it has.
`carbonvoice_send`	A text message into a conversation, as the agent's own Carbon Voice account, threaded under the message it answers.
`carbonvoice_channel`	One conversation: its name, its kind, and who is in it — how a sender id becomes a name.
`/carbonvoice`	Who the token is, and what came in: `status`, `recent [minutes]`.

Ask it "what did anyone say in Carbon Voice today?" and it calls carbonvoice_recent; ask it to answer someone and it threads the reply under that person's message.

Get it

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

Set it up

You need a Carbon Voice account — the agent speaks as that account — and a Personal Access Token from developer.carbonvoice.app.

export CARBONVOICE_PAT=cv_pat_...

The section names the variable, never the token:

"carbonvoice": {
  "pat_env": "CARBONVOICE_PAT",
  "base_url": "https://api.carbonvoice.app",
  "since_minutes": 60,
  "limit": 25,
  "exclude_self": true,
  "timeout": 20
}

Setting	What it does
`pat_env`	The variable holding the token.
`base_url`	The API to speak to; the default is Carbon Voice's own.
`since_minutes`	How far back `carbonvoice_recent` looks when the call does not say.
`limit`	How many messages it answers with at most.
`exclude_self`	Leave the agent's own messages out, so a reply is not read back as a new message.
`timeout`	How long one request may take, in seconds.

Use it

/carbonvoice says who the token is and what the window holds. The model reads the manual at (help :carbonvoice).

From eval, in the carbonvoice package:

(carbonvoice:recent :minutes 30)
(carbonvoice:read-message "<message id>")
(carbonvoice:send "<conversation id>" "the deploy is live" :reply-to "<message id>")
(carbonvoice:channel "<conversation id>")
(carbonvoice:whoami)

A voice memo arrives in two pieces: the audio, then its transcript a few seconds later. carbonvoice_recent marks a message it has no words for yet as pending, and carbonvoice_message answers the same message once Carbon Voice has transcribed it.

What it does not do

The original plugin is a chat platform: it keeps a Socket.IO connection, polls as a fallback, and answers people by itself as they speak, with an allowlist, approval prompts, acknowledgement reactions and a cursor file for the messages that arrived while it was down. Here the agent reads the inbox when it is asked and answers from the same account — the tools above — rather than a bot running beside the gateway. Attachments, voice replies through a text-to-speech pipeline and the /cv-allow-* commands are not carried over.