Tool Calling for LLM

LLM tool calling lets the PAL trigger functions based on what the user says during a conversation. Tools are reusable objects: create them once, and attach them to any number of PALs. They dispatch via the channel you pick on the delivery field - see Tool Delivery and Tool Authentication.

This page documents the tools registry (/v2/tools). If your PAL still defines tools inline under layers.llm.tools, see Legacy inline tool calling.

Tool Object

Top-Level Fields

Field	Type	Required	Description
`name`	string	✅	A unique identifier for the tool, scoped to your account. Must match OpenAI function naming rules (`^[a-zA-Z_][a-zA-Z0-9_]{0,63}$`).
`description`	string	✅	Natural language explanation of what the tool does. Used by the model to decide when to call it.
`parameters`	object	❌	JSON Schema object describing the tool’s input arguments. Defaults to `{}` (no arguments).
`origin`	string	❌	One of `llm`, `vision`, `audio`. Defaults to `llm`. This page covers `llm`; see Tool Calling for Perception for `vision` / `audio`.
`on_call`	string	❌	what the PAL does while the tool is running. See `on_call` reference. Defaults to `generate_filler` for `llm` tools; must be omitted for `vision` / `audio` tools.
`on_resolve`	string	❌	what the PAL does after the tool returns a result. See `on_resolve` reference. Defaults to `fire_and_forget`.
`static_filler`	string	❌	The exact line the PAL speaks while the tool runs when `on_call` is `static_filler`. Required in that case.
`delivery`	object	❌	How the tool call is delivered. See Tool Delivery. Defaults to `{"app_message": true}`.

`parameters`

Standard JSON Schema. Same shape as the OpenAI function.parameters object.

Field	Type	Required	Description
`type`	string	✅	Always `"object"`.
`properties`	object	✅	Map of parameter name to its schema (`type`, `description`, `enum`, etc.).
`required`	array of strings	❌	Names of mandatory parameters.

Examples

No arguments
Single required argument
Enum + optional argument

parameters

{
  "type": "object",
  "properties": {}
}

parameters

{
  "type": "object",
  "properties": {
    "city": {
      "type": "string",
      "description": "The city to get the weather for, e.g. San Francisco"
    }
  },
  "required": ["city"]
}

parameters

{
  "type": "object",
  "properties": {
    "city": {
      "type": "string",
      "description": "The city to get the weather for"
    },
    "unit": {
      "type": "string",
      "enum": ["celsius", "fahrenheit"],
      "description": "Temperature unit"
    }
  },
  "required": ["city"]
}

`on_call` reference

on_call controls what the PAL does while a tool is dispatched but has not yet returned. Defaults to generate_filler for llm tools; set it explicitly if you want a different behavior.

Value	Behavior
`generate_filler`	The LLM produces a short filler line itself (a `response_to_user` argument is injected into the tool’s schema). the PAL speaks it via TTS.
`static_filler`	The PAL speaks the configured `static_filler` string at dispatch time. The top-level `static_filler` field is required.
`silent`	The face stays silent while the tool runs.
`passthrough`	The tool spec is untouched. You manage the prompt and the schema. Use this when the LLM is already configured to preamble (i.e. the model naturally speaks a short line on its own before invoking a tool).

Examples

generate_filler
static_filler
silent
passthrough

{
  "on_call": "generate_filler"
}

The LLM picks the filler line per-call (e.g. “Let me check that for you.”).

{
  "on_call": "static_filler",
  "static_filler": "One moment - checking the weather now."
}

The PAL speaks the exact static_filler string every time the tool fires.

{
  "on_call": "silent"
}

The face says nothing while the tool runs.

{
  "on_call": "passthrough"
}

The tool spec is untouched. Use when your LLM is already configured to preamble before tool calls.

`on_resolve` reference

on_resolve controls what the PAL does with the tool’s result once it returns.

Value	Behavior
`generate_response`	The LLM regenerates a contextual reply using the tool’s result. Most common choice; The PAL speaks a natural-language summary of the result.
`response_in_result`	The tool’s response body is spoken verbatim via TTS - return `Content-Type: text/plain` with a single natural-language string. Use when the tool already returns a fully-formed natural-language reply. On timeout, non-2xx, or transport error, the tool call falls back to `generate_response` so the PAL still acknowledges the user.
`add_to_context`	The result lands silently in the conversation history as a system message. Nothing is spoken on the result-landing turn; the next user utterance’s LLM call sees it.
`fire_and_forget`	The result is not awaited or processed. Used when the tool’s side effect is what matters (e.g. logging a CRM event). The default when `on_resolve` is omitted.

Examples

generate_response
response_in_result
add_to_context
fire_and_forget

{
  "on_resolve": "generate_response"
}

The LLM regenerates a contextual reply using the tool’s result. Most common choice.

{
  "on_resolve": "response_in_result"
}

The tool’s response body is spoken verbatim via TTS. Return Content-Type: text/plain with a natural-language string.

{
  "on_resolve": "add_to_context"
}

The result lands silently in conversation history; the next user turn’s LLM call sees it.

{
  "on_resolve": "fire_and_forget"
}

The result is not awaited or processed. Use for side-effect-only tools (e.g. logging a CRM event).

Creating a Tool

Create tool

curl --request POST \
  --url https://tavusapi.com/v2/tools \
  --header 'Content-Type: application/json' \
  --header 'x-api-key: <api-key>' \
  --data '{
    "name": "get_current_weather",
    "description": "Get the current weather for a city.",
    "parameters": {
      "type": "object",
      "properties": {
        "city": {
          "type": "string",
          "description": "The city to get the weather for, e.g. San Francisco"
        },
        "unit": {
          "type": "string",
          "enum": ["celsius", "fahrenheit"]
        }
      },
      "required": ["city"]
    },
    "origin": "llm",
    "on_call": "generate_filler",
    "on_resolve": "generate_response",
    "delivery": {
      "api": {
        "url": "https://api.example.com/tools/get_weather",
        "method": "POST",
        "auth": { "type": "hmac", "secret": "whsec_long_random_string" },
        "timeout": 20
      }
    }
  }'

A successful create returns the full tool object including its tool_id:

Response

{
  "tool_id": "tabc123def456",
  "owner_id": 12345,
  "name": "get_current_weather",
  "description": "Get the current weather for a city.",
  "parameters": { "...": "..." },
  "delivery": { "api": { "...": "..." } },
  "is_system_tool": false,
  "origin": "llm",
  "on_call": "generate_filler",
  "on_resolve": "generate_response",
  "static_filler": null,
  "created_at": "2026-05-20 14:22:01.123456",
  "updated_at": "2026-05-20 14:22:01.123456"
}

The tool_id (prefix t…) is how you reference the tool when attaching it to a PAL or updating / deleting it.

Attaching Tools to a PAL

A PAL only sees a tool if it is attached. Attach one or more tools in a single call:

Attach tools to PAL

curl --request POST \
  --url https://tavusapi.com/v2/pals/{pal_id}/tools \
  --header 'Content-Type: application/json' \
  --header 'x-api-key: <api-key>' \
  --data '{
    "tool_ids": ["tabc123def456", "tdef456abc789"]
  }'

List the tools currently attached to a PAL:

List PAL tools

curl --request GET \
  --url https://tavusapi.com/v2/pals/{pal_id}/tools \
  --header 'x-api-key: <api-key>'

Detach a single tool from a PAL:

Detach tool from PAL

curl --request DELETE \
  --url https://tavusapi.com/v2/pals/{pal_id}/tools/{tool_id} \
  --header 'x-api-key: <api-key>'

Detaching removes the link between the PAL and the tool. The tool itself is untouched and still attached to any other PALs that use it.

End-to-End Example

Create the tool at /v2/tools (see above). Save the returned tool_id.
Attach it to a PAL at /v2/pals/{pal_id}/tools.
Start a conversation with that PAL. The tool is now available to the face.
Handle the tool call in your application:
- If delivery.app_message: true, listen for conversation.tool_call events in your frontend.
- If delivery.api is set, your endpoint receives the call. See Tool Delivery.
Return a result (only meaningful when on_resolve is not fire_and_forget):
- App message delivery: send a conversation.tool_result event with the matching tool_call_id. See Tool Delivery.
- API delivery: return a 2xx response with the result in the response body.

Errors

Status	When
`400`	Validation failure (bad `name`, missing `on_call` on an `llm` tool, both delivery channels set, invalid API URL, unknown enum value, etc.).
`400`	Attempted to attach a non-existent `tool_id`.
`409`	A tool with the same `name` already exists in your account. Names are unique per owner.
`404`	`tool_id` not found, or the authenticated account does not own it.

Replace <api-key> with your actual API key. You can generate one in the PAL Maker.

Getting started

Build

Deploy

Debug

Guides

Resources

Tool Calling for LLM

Tool Object

Top-Level Fields

`parameters`

Examples

`on_call` reference

Examples

`on_resolve` reference

Examples

Creating a Tool

Attaching Tools to a PAL

End-to-End Example

Errors

​Tool Object

​Top-Level Fields

​parameters

​Examples

​on_call reference

​Examples

​on_resolve reference

​Examples

​Creating a Tool

​Attaching Tools to a PAL

​End-to-End Example

​Errors

Tool Object

Top-Level Fields

`parameters`

Examples

`on_call` reference

Examples

`on_resolve` reference

Examples

Creating a Tool

Attaching Tools to a PAL

End-to-End Example

Errors