The agent's settings

Part of the MCP server: every tool of this stage, what each parameter takes — read off the tool's own schema, as the assistant sees it — and one real call, answered by Pinecall Cloud.

agent#

The agent's settings — voice, models, language, greeting, memory policy, knowledge by heart: read, set, cleared, history, diff, rollback.

agent reads and writes the agent's settings in the sandbox — the voice, the models, the language, the greeting, the memory policy, what it knows by heart — each change a new version, changed without a deploy. They are your own settings unless team writes the team's. history, diff and rollback read and undo versions. A model is named as vendor/model, a vendor, a model, or a tier (haiku). greeting is the words, improvise for the model's own opening, or improvise:… with an instruction; hangup is when, in words, or any; end-of-turn is who ends the caller's turn. temperature is the model's; llm-builds and llm-options (and the stt, tts and judge twins) are a class of the vendor's plugin and its keyword arguments, run only on the org's own key for that vendor. judge is the model this agent's calls are judged on, named as llm is, over the org's (judging); on the org's own key for the vendor — a local model's server through judge-options {"base_url": …} — its evals are never billed. A field the class declares itself is listed in fixed, and setting it is refused naming the class.

{
  "action": "set",
  "settings": {
    "greeting": "Thanks for calling. How can I help?",
    "endpointing-ms": 500
  },
  "note": "from the docs"
}

{
  "world": "sandbox",
  "yours": {
    "holder": "m_4f1c2a9b0e7d",
    "version": 26,
    "author": "m_4f1c2a9b0e7d",
    "note": "from the docs",
    "set_at": 1791490523.58,
    "config": {
      "greeting": {
        "say": "Thanks for calling. How can I help?"
      },
      "turn": {
        "endpointing_ms": 500
      }
    }
  },
  "team": null,
  "production": null
}

lexicon#

The agent's words: how its voice says one (add), the words its ears must catch (hear), taking them out (rm), and its versions.

lexicon fixes words the voice says wrong (add, with how to say it) and names the words the ears must catch (hear): a brand, a doctor's name, a street. Each change is a version, kept without a deploy.

{
  "action": "add",
  "word": "Pinecall",
  "say": "pine call"
}

{
  "world": "sandbox",
  "yours": {
    "holder": "m_4f1c2a9b0e7d",
    "version": 8,
    "author": "m_4f1c2a9b0e7d",
    "note": null,
    "set_at": 1791490532.24,
    "lexicon": {
      "said": [
        {
          "word": "Pinecall",
          "spoken": "pine call"
        }
      ],
      "heard": []
    }
  },
  "team": null,
  "production": null
}

voices#

A voice vendor's voices in a language: the id an agent's voice setting takes, the name, gender and accent.

voices lists the voices to choose from, the id first; set one with agent (settings.voice, and settings.tts for its vendor). voice_sample lets the person hear one first.

{
  "language": "es",
  "country": "ES"
}

{
  "voices": [
    {
      "id": "a7beff01-8f8b-4809-bfe6-e2166e57e0c2",
      "name": "Iria - Thoughtful Communicator",
      "language": "es",
      "description": "Her measured and thoughtful delivery creates an engaging experience, perfect for detailed explanations and educational content.",
      "gender": "feminine",
      "country": "ES",
      "accent": "castilian"
    },
    {
      "id": "ad38904c-0ce9-42b1-9159-5ad5352ef089",
      "name": "Celia - Practical Analyst",
      "language": "es",
      "description": "With a thoughtful and analytical tone, this voice simplifies issues and provides actionable insights.",
      "gender": "feminine",
      "country": "ES",
      "accent": "castilian"
    },
    {
      "id": "de38f545-c574-44e8-9b54-a7d6fec1c6b1",
      "name": "Marta - Friendly Guide",
      "language": "es",
      "description": "Approachable Spanish female ideal for customer care and support.",
      "gender": "feminine",
      "country": "ES",
      "accent": "castilian"
    },
    "… 5 more"
  ]
}

voice_sample#

Have the gateway say a line in a voice, as a call would, and save it as a WAV the person can play; with how long the vendor took.

voice_sample says a line through the vendor's own plugin and saves the WAV under the project's .pinecall/voices/: this server plays nothing, so the answer is the file's path, for the person to open, with the first-audio and whole-sentence times.

{
  "voice": "a7beff01-8f8b-4809-bfe6-e2166e57e0c2",
  "text": "Hola, gracias por llamar."
}

{
  "file": "/Users/you/front-desk/.pinecall/voices/a7beff01-8f8b-4809-bfe6-e2166e57e0c2.wav",
  "tts": "cartesia",
  "first_audio_ms": 418,
  "whole_sentence_ms": 855
}