> ## Documentation Index
> Fetch the complete documentation index at: https://docs.luly.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Create a Web Agent

> Configure all of the settings for a new web agent.

### Headers

<ParamField header="authorization" type="string" required={true}>
  Your API key for authentication.
</ParamField>

### Body

<ParamField body="prompt" type="string" required>
  Provide instructions, relevant information, and examples of the ideal conversation flow.

  <Accordion title="Best Practices">
    #### Out-of-the-Box Behaviors (Summarized):

    * Speech pattern: Direct, concise, casual
    * Spells out symbols, acronyms, abbreviations, percentages, etc. (\$4,000,000 -> “four million dollars”)
    * Asks clarifying questions
    * Ends call when objective is complete or voicemail is detected

    #### Prompting Tips:

    * Want to easily test out exactly how your agent will behave?
      * [Try out Agent Testing!](https://portal.luly.ai/login)
    * Aim for less than >2,000 characters where possible.
    * Simple, direct prompts are the most predictable and reliable.
    * Frame instructions positively:
      * `"Do this"` rather than `"Don't do this"`.
      * Ex. “Keep the conversation casual” rather than “Don’t be too formal”.
      * This gives concrete examples of what to do, instead of leaving expected behavior open to interpretation.
  </Accordion>
</ParamField>

<ParamField body="voice" type="string">
  Set your agent’s voice - all available voices can be found with the [List Voices](https://docs.luly.ai/api-v1/get/voices) endpoint.
</ParamField>

<ParamField body="analysis_schema" type="object">
  Define a JSON schema for how you want to get information about the call - information like email addresses, names, appointment times or any other type of custom data.

  In the webhook response or whenever you retrieve call data later, you’ll get the data you defined back under analysis.

  For example, if you wanted to retrieve this information from the call:

  ```json
  "analysis_schema": {
  "email_address": "email",
  "first_name": "string",
  "last_name": "string",
  "wants_to_book_appointment": "boolean",
  "appointment_time": "YYYY-MM-DD HH:MM:SS"
  }
  ```

  You would get it filled out like this in your webhook once the call completes:

  ```json
  "analysis": {
  "email_address": "johndoe@gmail.com",
  "first_name": "John",
  "last_name": "Doe",
  "wants_to_book_appointment": true,
  "appointment_time": "2024-01-01 12:00:00"
  }

  ```
</ParamField>

<ParamField body="metadata" type="object">
  Add any additional information you want to associate with the call. This can be useful for tracking or categorizing calls.
</ParamField>

<ParamField body="pathway_id" type="string">
  Set the pathway that your agent will follow. This will override the `prompt` field, so there is no need to pass the ‘prompt’ field if you are setting a pathway.

  Warning: Setting a pathway will set the following fields to `null` / their default value - `prompt`, `first_sentence`, `model`, `dynamic_data`, `tools`, `transfer_list`

  Set to `null` or an empty string to clear the pathway.
</ParamField>

<ParamField body="language" type="string" default="ENG">
  Select a supported language of your choice. Optimizes every part of our API for that language - transcription, speech, and other inner workings.

  Supported Languages and their codes:

  * English: `ENG`
  * Spanish: `ESP`
  * French: `FRE`
  * Polish: `POL`
  * German: `GER`
  * Italian: `ITA`
  * Brazilian Portuguese: `PBR`
  * Portuguese: `POR`
</ParamField>

<ParamField body="model" type="string" default="enhanced">
  Select a model to use for your call.

  Options: `base`, `turbo` and `enhanced`.

  In nearly all cases, `enhanced` is the best choice for now.

  <Accordion title="Model differences">
    There are three different ways to use luly:

    * `model: base`
      * The original, follows scripts/procedures most effectively.
      * Supports all features and capabilities.
      * Best for Custom Tools
    * `model: enhanced`
      * Much faster latency and very conversational, works best with objective-based prompts.
      * Supports all features and capabilities.
    * `model: turbo`
      * The absolute fastest latency possible, can be verbose at times
      * Limited capabilities currently (excludes Transferring, IVR navigation, Custom Tools)
      * Extremely realistic conversation capabilities
  </Accordion>
</ParamField>

<ParamField body="first_sentence" type="string">
  A phrase that your call will start with instead of a generating one on the fly. This works both with and without `wait_for_greeting`. Can be more than one sentence, but must be less than 200 characters.

  To remove, set to null or an empty string.
</ParamField>

<ParamField body="tools" type="array">
  Interact with the real world through API calls.

  Detailed tutorial here: [Custom Tools](https://docs.luly.ai/tutorials/custom-tools)
</ParamField>

<ParamField body="dynamic_data" type="object">
  Integrate data from external APIs into your agent’s knowledge.

  Set to null or an empty string to clear dynamic data settings.

  Detailed usage in the [Send Call](https://docs.luly.ai/api-v1/post/calls) endpoint.
</ParamField>

<ParamField body="interruption_threshold" type="number" default="100">
  Adjusts how patient the AI is when waiting for the user to finish speaking.

  Lower values mean the AI will respond more quickly, while higher values mean the AI will wait longer before responding.

  Recommended range: 50-200

  * 50: Extremely quick, back and forth conversation
  * 100: Balanced to respond at a natural pace
  * 200: Very patient, allows for long pauses and interruptions. Ideal for collecting detailed information.

  Try to start with 100 and make small adjustments in increments of \~10 as needed for your use case.
</ParamField>

<ParamField body="keywords" type="string[]" default="[]">
  These words will be boosted in the transcription engine - recommended for proper nouns or words that are frequently mis-transcribed.

  For example, if the word “Reece” is frequently transcribed as a homonym like “Reese” you could do this:

  ```json
      {
          "keywords": ["Reece"]
      }
  ```

  For stronger keyword boosts, you can place a colon then a boost factor after the word. The default boost factor is 2.

  ```json
      {
          "keywords": ["Reece:3"]
      }

  ```
</ParamField>

<ParamField body="max_duration" type="integer (minutes)" default="30">
  The maximum duration that calls to your agent can last before being automatically terminated.

  Set to `null` to reset to default.
</ParamField>

### Response

<ResponseField name="status" type="string">
  Can be `success` or `error`.
</ResponseField>

<ResponseField name="call_id" type="string">
  A unique identifier for the call (present only if status is `success`).
</ResponseField>

<ResponseExample>
  ```json Response
  {
    "agent": {
          "agent_id": "2c565dc7-f41f-43db-a99f-e4c8ba9d7d18",
          "dynamic_data": null,
          "interruption_threshold": null,
          "first_sentence": null,
          "model": "enhanced",
          "voice_settings": null,
          "voice": "maya",
          "prompt": "...",
          "temperature": null,
          "max_duration": 30,
          "language": "ENG",
          "tools": null
    }
  }
  ```
</ResponseExample>