feat(agentkit): add support for mllm & tts

Author: seymourtang

docs/concepts/vendors.md

@@ -70,6 +70,7 @@ Used with `agent.with_tts()`. Each TTS vendor produces audio at a specific sampl
 | `DeepgramTTS` | Deepgram | `api_key`, `model` | Configurable |
 | `GradiumTTS` | Gradium | `api_key` | Configurable |
 | `MistralTTS` | Mistral | `api_key` | — |
+| `TypecastTTS` | Typecast | `api_key`, `voice_id`, `model` | — |
 | `SarvamTTS` | Sarvam | `api_key` | — |
 | `XaiTTS` | xAI | `api_key`, `language` | Configurable |
 
@@ -109,6 +110,8 @@ Used with `agent.with_stt()`.
 
 Use `turn_detection.language` for Agora interaction language; it defaults to `en-US`. STT vendor `language` options are serialized under `asr.params` using each provider's own format. If `with_stt()` is omitted, AgentKit defaults to `AresSTT` for global clients and `FengmingSTT` for `Area.CN` clients. Ares does not take a provider language option; AgentKit uses `turn_detection.language` for REST `asr.language`.
 
+`AresSTT` and `FengmingSTT` accept an optional `keywords: List[str]` hotword list.
+
 | Class | Provider | Required Parameters |
 |---|---|---|
 | `SpeechmaticsSTT` | Speechmatics | `api_key`, `language` |
@@ -146,12 +149,14 @@ stt = DeepgramSTT(api_key='your-deepgram-key', language='en-US', model='nova-2')
 
 Used with `agent.with_mllm()` for the [MLLM flow](../guides/mllm-flow.md). These handle audio input and output end-to-end.
 
-| Class | Provider | Required Parameters |
-|---|---|---|
-| `OpenAIRealtime` | OpenAI Realtime | `api_key`; optional `turn_detection` |
-| `GeminiLive` | Google Gemini Live API | `api_key`, `model`; optional `turn_detection` |
-| `VertexAI` | Vertex AI (Gemini Live) | `model`, `project_id`, `location`, `adc_credentials_string`; optional `turn_detection` |
-| `XaiGrok` | xAI Grok (`mllm.vendor`: `xai`) | `api_key`; optional `voice`, `language`, `sample_rate`, `turn_detection` |
+| Class | Provider | Area | Required Parameters |
+|---|---|---|---|
+| `OpenAIRealtime` | OpenAI Realtime | Global | `api_key`; optional `turn_detection` |
+| `AzureOpenAIRealtime` | Azure OpenAI Realtime | Global | `api_key`, `url`, `turn_detection`; optional `max_history` |
+| `GeminiLive` | Google Gemini Live API | Global | `api_key`, `model`; optional `turn_detection` |
+| `VertexAI` | Vertex AI (Gemini Live) | Global | `model`, `project_id`, `location`, `adc_credentials_string`; optional `turn_detection` |
+| `XaiGrok` | xAI Grok (`mllm.vendor`: `xai`) | Global | `api_key`; optional `voice`, `language`, `sample_rate`, `turn_detection` |
+| `QwenOmni` | Alibaba Cloud Qwen Omni Realtime | CN | `api_key`, `turn_detection`; optional `url` |
 
 <!-- snippet: executable -->
 ```python
@@ -160,6 +165,12 @@ from agora_agent import OpenAIRealtime
 mllm = OpenAIRealtime(api_key='your-openai-key', model='gpt-4o-realtime-preview')
 ```
 
+Import the CN-only Qwen helper from `agora_agent.cn`:
+
+```python
+from agora_agent.cn import QwenOmni
+```
+
 ## Avatar Vendors
 
 Used with `agent.with_avatar()` in the cascading ASR + LLM + TTS pipeline. Some avatars require specific TTS sample rates — see [Avatar Integration](../guides/avatars.md).

docs/guides/mllm-flow.md

@@ -1,7 +1,7 @@
 ---
 sidebar_position: 2
 title: MLLM Flow (Multimodal)
-description: Use OpenAI Realtime, Gemini Live, Vertex AI, or xAI for end-to-end audio processing.
+description: Use supported global or CN realtime multimodal models for end-to-end audio processing.
 ---
 
 # MLLM Flow (Multimodal)
@@ -11,9 +11,11 @@ The MLLM (Multimodal LLM) flow uses a single model to handle both audio input an
 MLLM vendors supported by AgentKit:
 
 - **OpenAI Realtime** — `gpt-4o-realtime-preview` and related models
+- **Azure OpenAI Realtime** — global Azure-hosted realtime deployments
 - **Gemini Live** — direct Google AI API access for audio-native Gemini models
 - **Vertex AI** — Gemini Live through Google Cloud Vertex AI
 - **xAI Grok** — xAI Realtime API
+- **Qwen Omni** — Alibaba Cloud Qwen Omni Realtime for `Area.CN`
 
 ## Enable MLLM Mode
 
@@ -167,7 +169,7 @@ mllm = OpenAIRealtime(
 |---|---|---|
 | Latency | Lower — single model, no pipeline | Higher — three models in sequence |
 | Voice control | Model-dependent | Full vendor choice for TTS |
-| Vendor flexibility | Limited to supported MLLM providers (OpenAI Realtime, Gemini Live, Vertex AI, xAI Grok) | Mix and match LLM, TTS, and STT vendors |
+| Vendor flexibility | Limited to supported global and CN MLLM providers | Mix and match LLM, TTS, and STT vendors |
 | Audio understanding | Model hears tone, pacing, emotion | STT produces text only |
 
 ## Next Steps

docs/guides/regional-routing.md

@@ -38,10 +38,10 @@ client = Agora(
 Bind `client` into `Agent(client=client, ...)` and construct vendors directly with SDK classes. The bound client selects `CNAgent` or `GlobalAgent` for IDE hints based on `area`, but does not restrict which vendor classes you can configure.
 If you omit `with_stt()`, AgentKit uses `FengmingSTT` by default for `Area.CN` clients and `AresSTT` for global clients.
 
-| Client area | STT classes | LLM classes | TTS classes | Avatar classes |
-|---|---|---|---|---|
-| `Area.US`, `Area.EU`, `Area.AP` | `DeepgramSTT`, `SpeechmaticsSTT`, `MicrosoftSTT`, `OpenAISTT`, `GoogleSTT`, `AmazonSTT`, `AssemblyAISTT`, `AresSTT`, `SarvamSTT`, `XaiSTT` | `OpenAI`, `AzureOpenAI`, `Anthropic`, `Gemini`, `Groq`, `VertexAILLM`, `AmazonBedrock`, `Dify`, `CustomLLM` | `ElevenLabsTTS`, `MicrosoftTTS`, `OpenAITTS`, `CartesiaTTS`, `GoogleTTS`, `AmazonTTS`, `DeepgramTTS`, `GradiumTTS`, `MistralTTS`, `HumeAITTS`, `RimeTTS`, `FishAudioTTS`, `MiniMaxTTS`, `MurfTTS`, `SarvamTTS`, `GenericTTS`, `XaiTTS` | `LiveAvatarAvatar`, `HeyGenAvatar`, `AkoolAvatar`, `AnamAvatar`, `GenericAvatar` |
-| `Area.CN` | `FengmingSTT`, `TencentSTT`, `MicrosoftCNSTT`, `XfyunSTT`, `XfyunBigModelSTT`, `XfyunDialectSTT` | `AliyunLLM`, `BytedanceLLM`, `DeepSeekLLM`, `TencentLLM` | `MiniMaxCNTTS`, `TencentTTS`, `BytedanceTTS`, `MicrosoftCNTTS`, `CosyVoiceTTS`, `BytedanceDuplexTTS`, `StepFunTTS`, `GenericTTS` | `SenseTimeAvatar`, `SpatiusAvatar` |
+| Client area | STT classes | LLM classes | MLLM classes | TTS classes | Avatar classes |
+|---|---|---|---|---|---|
+| `Area.US`, `Area.EU`, `Area.AP` | `DeepgramSTT`, `SpeechmaticsSTT`, `MicrosoftSTT`, `OpenAISTT`, `GoogleSTT`, `AmazonSTT`, `AssemblyAISTT`, `AresSTT`, `SarvamSTT`, `XaiSTT` | `OpenAI`, `AzureOpenAI`, `Anthropic`, `Gemini`, `Groq`, `VertexAILLM`, `AmazonBedrock`, `Dify`, `CustomLLM` | `OpenAIRealtime`, `AzureOpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok` | `ElevenLabsTTS`, `MicrosoftTTS`, `OpenAITTS`, `CartesiaTTS`, `GoogleTTS`, `AmazonTTS`, `DeepgramTTS`, `GradiumTTS`, `MistralTTS`, `TypecastTTS`, `HumeAITTS`, `RimeTTS`, `FishAudioTTS`, `MiniMaxTTS`, `MurfTTS`, `SarvamTTS`, `GenericTTS`, `XaiTTS` | `LiveAvatarAvatar`, `HeyGenAvatar`, `AkoolAvatar`, `AnamAvatar`, `GenericAvatar` |
+| `Area.CN` | `FengmingSTT`, `TencentSTT`, `MicrosoftCNSTT`, `XfyunSTT`, `XfyunBigModelSTT`, `XfyunDialectSTT` | `AliyunLLM`, `BytedanceLLM`, `DeepSeekLLM`, `TencentLLM` | `QwenOmni` | `MiniMaxCNTTS`, `TencentTTS`, `BytedanceTTS`, `MicrosoftCNTTS`, `CosyVoiceTTS`, `BytedanceDuplexTTS`, `StepFunTTS`, `GenericTTS` | `SenseTimeAvatar`, `SpatiusAvatar` |
 
 Global client example:
 

docs/index.md

@@ -17,12 +17,12 @@ The Agora Conversational AI Python SDK lets you build voice-powered AI agents on
 
 **Cascading flow** uses ASR -> LLM -> TTS and supports the broadest set of vendor combinations.
 
-**MLLM flow** uses a multimodal model such as OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok for end-to-end audio.
+**MLLM flow** uses a multimodal model such as OpenAI Realtime, Azure OpenAI Realtime, Gemini Live, Vertex AI, xAI Grok, or Qwen Omni for end-to-end audio.
 
 ## Start here
 
 - Start with [Quick Start](./getting-started/quick-start.md). It shows the baseline app-credentials setup and starts a cascading ASR -> LLM -> TTS agent.
-- Use [MLLM Flow](./guides/mllm-flow.md) when your agent uses one realtime multimodal model, such as OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok.
+- Use [MLLM Flow](./guides/mllm-flow.md) when your agent uses one supported realtime multimodal model.
 - Use [Cascading Flow](./guides/cascading-flow.md) for more examples of the default ASR -> LLM -> TTS flow, including provider-specific configuration.
 
 ## How the SDK is organized
@@ -46,7 +46,7 @@ The Agora Conversational AI Python SDK lets you build voice-powered AI agents on
 | [AgentSession](./concepts/session.md) | Manage the agent lifecycle |
 | [Vendors](./concepts/vendors.md) | Browse all LLM, TTS, STT, MLLM, and Avatar providers |
 | [Cascading Flow](./guides/cascading-flow.md) | Build an ASR -> LLM -> TTS pipeline |
-| [MLLM Flow](./guides/mllm-flow.md) | Use OpenAI Realtime, Gemini Live, Vertex AI, or xAI Grok for end-to-end audio |
+| [MLLM Flow](./guides/mllm-flow.md) | Use a global or CN realtime multimodal model for end-to-end audio |
 | [Avatars](./guides/avatars.md) | Add a digital avatar with LiveAvatar, Akool, Anam, Generic Avatar, SenseTime, or Spatius |
 | [Regional Routing](./guides/regional-routing.md) | Route requests to the nearest region |
 | [Error Handling](./guides/error-handling.md) | Handle API errors with ApiError |

docs/reference/vendors.md

@@ -10,17 +10,18 @@ All vendor classes are available from `agora_agent`:
 
 <!-- snippet: fragment -->
 ```python
-from agora_agent import OpenAI, ElevenLabsTTS, DeepgramTTS, DeepgramSTT, OpenAIRealtime, XaiGrok, GenericAvatar
+from agora_agent import AzureOpenAIRealtime, DeepgramSTT, OpenAI, TypecastTTS
+from agora_agent.cn import QwenOmni
 ```
 
 ## Recommended vendors by area
 
 Construct vendors directly from `agora_agent`, then bind a client with `Agent(client=client, ...)`. The bound client selects `CNAgent` or `GlobalAgent` for IDE hints based on `area`, but does not restrict which vendor classes you can configure.
 
-| Area | STT classes | LLM classes | TTS classes | Avatar classes |
-|---|---|---|---|---|
-| `Area.US`, `Area.EU`, `Area.AP` | `DeepgramSTT`, `SpeechmaticsSTT`, `MicrosoftSTT`, `OpenAISTT`, `GoogleSTT`, `AmazonSTT`, `AssemblyAISTT`, `AresSTT`, `SarvamSTT`, `XaiSTT` | `OpenAI`, `AzureOpenAI`, `Anthropic`, `Gemini`, `Groq`, `VertexAILLM`, `AmazonBedrock`, `Dify`, `CustomLLM` | `ElevenLabsTTS`, `MicrosoftTTS`, `OpenAITTS`, `CartesiaTTS`, `GoogleTTS`, `AmazonTTS`, `DeepgramTTS`, `GradiumTTS`, `MistralTTS`, `HumeAITTS`, `RimeTTS`, `FishAudioTTS`, `MiniMaxTTS`, `MurfTTS`, `SarvamTTS`, `GenericTTS`, `XaiTTS` | `LiveAvatarAvatar`, `HeyGenAvatar`, `AkoolAvatar`, `AnamAvatar`, `GenericAvatar` |
-| `Area.CN` | `FengmingSTT`, `TencentSTT`, `MicrosoftCNSTT`, `XfyunSTT`, `XfyunBigModelSTT`, `XfyunDialectSTT` | `AliyunLLM`, `BytedanceLLM`, `DeepSeekLLM`, `TencentLLM` | `MiniMaxCNTTS`, `TencentTTS`, `BytedanceTTS`, `MicrosoftCNTTS`, `CosyVoiceTTS`, `BytedanceDuplexTTS`, `StepFunTTS`, `GenericTTS` | `SenseTimeAvatar`, `SpatiusAvatar` |
+| Area | STT classes | LLM classes | MLLM classes | TTS classes | Avatar classes |
+|---|---|---|---|---|---|
+| `Area.US`, `Area.EU`, `Area.AP` | `DeepgramSTT`, `SpeechmaticsSTT`, `MicrosoftSTT`, `OpenAISTT`, `GoogleSTT`, `AmazonSTT`, `AssemblyAISTT`, `AresSTT`, `SarvamSTT`, `XaiSTT` | `OpenAI`, `AzureOpenAI`, `Anthropic`, `Gemini`, `Groq`, `VertexAILLM`, `AmazonBedrock`, `Dify`, `CustomLLM` | `OpenAIRealtime`, `AzureOpenAIRealtime`, `GeminiLive`, `VertexAI`, `XaiGrok` | `ElevenLabsTTS`, `MicrosoftTTS`, `OpenAITTS`, `CartesiaTTS`, `GoogleTTS`, `AmazonTTS`, `DeepgramTTS`, `GradiumTTS`, `MistralTTS`, `TypecastTTS`, `HumeAITTS`, `RimeTTS`, `FishAudioTTS`, `MiniMaxTTS`, `MurfTTS`, `SarvamTTS`, `GenericTTS`, `XaiTTS` | `LiveAvatarAvatar`, `HeyGenAvatar`, `AkoolAvatar`, `AnamAvatar`, `GenericAvatar` |
+| `Area.CN` | `FengmingSTT`, `TencentSTT`, `MicrosoftCNSTT`, `XfyunSTT`, `XfyunBigModelSTT`, `XfyunDialectSTT` | `AliyunLLM`, `BytedanceLLM`, `DeepSeekLLM`, `TencentLLM` | `QwenOmni` | `MiniMaxCNTTS`, `TencentTTS`, `BytedanceTTS`, `MicrosoftCNTTS`, `CosyVoiceTTS`, `BytedanceDuplexTTS`, `StepFunTTS`, `GenericTTS` | `SenseTimeAvatar`, `SpatiusAvatar` |
 
 Global example:
 
@@ -411,6 +412,16 @@ AgentKit serializes `credential_mode` at the top level of the Rime TTS configura
 | `additional_params` | `Dict[str, Any]` | No | `None` | Additional Mistral TTS parameters, flattened into `params` |
 | `skip_patterns` | `List[int]` | No | `None` | Skip patterns |
 
+### `TypecastTTS`
+
+| Parameter | Type | Required | Default | Description |
+|---|---|---|---|---|
+| `api_key` | `str` | Yes | — | Typecast API key |
+| `voice_id` | `str` | Yes | — | Typecast voice identifier |
+| `model` | `str` | Yes | — | Typecast TTS model name, for example `ssfm-v30` |
+| `additional_params` | `Dict[str, Any]` | No | `None` | Additional Typecast parameters, flattened into `params` |
+| `skip_patterns` | `List[int]` | No | `None` | Skip patterns |
+
 ### `MurfTTS`
 
 | Parameter | Type | Required | Default | Description |
@@ -550,6 +561,7 @@ For `nova-2` and `nova-3`, omit `api_key` to use Agora-managed credentials. For
 
 | Parameter | Type | Required | Default | Description |
 |---|---|---|---|---|
+| `keywords` | `List[str]` | No | `None` | Hotwords that improve ASR accuracy |
 | `additional_params` | `Dict[str, Any]` | No | `None` | Additional parameters |
 
 ### `SarvamSTT`
@@ -700,7 +712,9 @@ All CN TTS vendor classes support `skip_patterns` and `additional_params`.
 
 #### `FengmingSTT`
 
-No constructor parameters. Use `FengmingSTT()`.
+| Parameter | Type | Required | Default | Description |
+|---|---|---|---|---|
+| `keywords` | `List[str]` | No | `None` | Hotwords that improve ASR accuracy |
 
 #### `XfyunSTT`
 
@@ -789,6 +803,46 @@ No constructor parameters. Use `FengmingSTT()`.
 | `params` | `Dict[str, Any]` | No | `None` | Additional parameters |
 | `turn_detection` | `MllmTurnDetectionConfig` | No | `None` | MLLM turn detection configuration; overrides top-level `turn_detection` |
 
+### `AzureOpenAIRealtime`
+
+Global Azure OpenAI Realtime vendor (`mllm.vendor`: `"azure"`).
+
+| Parameter | Type | Required | Default | Description |
+|---|---|---|---|---|
+| `api_key` | `str` | Yes | — | Azure OpenAI API key |
+| `url` | `str` | Yes | — | Azure OpenAI Realtime WebSocket URL |
+| `model` | `str` | No | `None` | Model or deployment name |
+| `voice` | `str` | No | `None` | Voice identifier |
+| `instructions` | `str` | No | `None` | System instructions |
+| `input_audio_transcription` | `Dict[str, Any]` | No | `None` | Audio transcription settings |
+| `max_history` | `int` | No | `None` | Number of conversation history messages to cache; Azure-only MLLM field |
+| `greeting_message` | `str` | No | `None` | Greeting message |
+| `failure_message` | `str` | No | `None` | Message played when the model call fails |
+| `output_modalities` | `List[str]` | No | `None` | Output modalities |
+| `messages` | `List[Dict]` | No | `None` | Conversation messages |
+| `params` | `Dict[str, Any]` | No | `None` | Additional Azure OpenAI parameters |
+| `turn_detection` | `MllmTurnDetectionConfig` | Yes | — | Required MLLM turn detection configuration; overrides top-level `turn_detection` |
+
+### `QwenOmni`
+
+CN Alibaba Cloud Qwen Omni Realtime vendor (`mllm.vendor`: `"qwen_omni"`). Import it with `from agora_agent.cn import QwenOmni`.
+
+| Parameter | Type | Required | Default | Description |
+|---|---|---|---|---|
+| `api_key` | `str` | Yes | — | Alibaba Cloud Model Studio API key |
+| `url` | `str` | No | `None` | Qwen Omni Realtime WebSocket URL |
+| `model` | `str` | No | `None` | Qwen Omni Realtime model name |
+| `voice` | `str` | No | `None` | Voice identifier |
+| `instructions` | `str` | No | `None` | System instructions |
+| `input_audio_transcription` | `Dict[str, Any]` | No | `None` | Audio transcription settings |
+| `greeting_message` | `str` | No | `None` | Greeting message |
+| `failure_message` | `str` | No | `None` | Message played when the model call fails |
+| `input_modalities` | `List[str]` | No | `None` | Input modalities |
+| `output_modalities` | `List[str]` | No | `None` | Output modalities |
+| `messages` | `List[Dict]` | No | `None` | Conversation messages |
+| `params` | `Dict[str, Any]` | No | `None` | Additional Qwen Omni parameters |
+| `turn_detection` | `MllmTurnDetectionConfig` | Yes | — | Required MLLM turn detection configuration; overrides top-level `turn_detection` |
+
 ### `GeminiLive`
 
 | Parameter | Type | Required | Default | Description |

src/agora_agent/agentkit/__init__.py

@@ -156,6 +156,7 @@
     AresSTT,
     AssemblyAISTT,
     AzureOpenAI,
+    AzureOpenAIRealtime,
     BaseAvatar,
     BaseLLM,
     BaseMLLM,
@@ -195,9 +196,11 @@
     OpenAITTS,
     GenericTTS,
     RimeTTS,
+    QwenOmni,
     SampleRate,
     SarvamSTT,
     SarvamTTS,
+    TypecastTTS,
     SpeechmaticsSTT,
     XaiSTT,
     XaiTTS,
@@ -403,6 +406,7 @@
     "MiniMaxTTS",
     "MiniMaxCNTTS",
     "MistralTTS",
+    "TypecastTTS",
     "TencentTTS",
     "BytedanceTTS",
     "CosyVoiceTTS",
@@ -428,7 +432,9 @@
     "XfyunDialectSTT",
     "XfyunSTT",
     "OpenAIRealtime",
+    "AzureOpenAIRealtime",
     "GeminiLive",
+    "QwenOmni",
     "VertexAI",
     "XaiGrok",
     "HeyGenAvatar",

src/agora_agent/agentkit/regional_agent.py

@@ -2,7 +2,7 @@
 
 import typing
 from .agent import Agent
-from .vendors.base import BaseAvatar, BaseLLM, BaseSTT, BaseTTS
+from .vendors.base import BaseAvatar, BaseLLM, BaseMLLM, BaseSTT, BaseTTS
 from .vendors.cn import (
     AliyunLLM,
     BytedanceDuplexTTS,
@@ -14,6 +14,7 @@
     MiniMaxTTS as MiniMaxCNTTS,
     MicrosoftSTT as MicrosoftCNSTT,
     MicrosoftTTS as MicrosoftCNTTS,
+    QwenOmni,
     SenseTimeAvatar,
     SpatiusAvatar,
     StepFunTTS,
@@ -47,6 +48,7 @@
     OpenAI,
     VertexAILLM,
 )
+from .vendors.mllm import AzureOpenAIRealtime, GeminiLive, OpenAIRealtime, VertexAI, XaiGrok
 from .vendors.tts import (
     AmazonTTS,
     CartesiaTTS,
@@ -64,13 +66,15 @@
     GenericTTS,
     RimeTTS,
     SarvamTTS,
+    TypecastTTS,
     XaiTTS,
 )
 from .vendors.avatar import AkoolAvatar, AnamAvatar, GenericAvatar, HeyGenAvatar, LiveAvatarAvatar
 
 CNSTT = typing.Union[TencentSTT, FengmingSTT, MicrosoftCNSTT, XfyunSTT, XfyunBigModelSTT, XfyunDialectSTT]
 CNTTS = typing.Union[MiniMaxCNTTS, TencentTTS, BytedanceTTS, MicrosoftCNTTS, CosyVoiceTTS, BytedanceDuplexTTS, StepFunTTS, GenericTTS]
 CNLLM = typing.Union[AliyunLLM, BytedanceLLM, DeepSeekLLM, TencentLLM]
+CNMLLM = QwenOmni
 CNAvatar = typing.Union[SenseTimeAvatar, SpatiusAvatar]
 
 GlobalSTT = typing.Union[
@@ -103,6 +107,7 @@
     DeepgramTTS,
     GradiumTTS,
     MistralTTS,
+    TypecastTTS,
 ]
 GlobalLLM = typing.Union[
     OpenAI,
@@ -115,6 +120,7 @@
     Dify,
     CustomLLM,
 ]
+GlobalMLLM = typing.Union[OpenAIRealtime, AzureOpenAIRealtime, GeminiLive, VertexAI, XaiGrok]
 GlobalAvatar = typing.Union[AkoolAvatar, LiveAvatarAvatar, AnamAvatar, GenericAvatar, HeyGenAvatar]
 
 
@@ -128,6 +134,9 @@ def with_llm(self, vendor: BaseLLM) -> "CNAgent":
     def with_tts(self, vendor: BaseTTS) -> "CNAgent":
         return typing.cast("CNAgent", super().with_tts(vendor))
 
+    def with_mllm(self, vendor: BaseMLLM) -> "CNAgent":
+        return typing.cast("CNAgent", super().with_mllm(vendor))
+
     def with_avatar(self, vendor: BaseAvatar) -> "CNAgent":
         return typing.cast("CNAgent", super().with_avatar(vendor))
 
@@ -142,6 +151,9 @@ def with_llm(self, vendor: BaseLLM) -> "GlobalAgent":
     def with_tts(self, vendor: BaseTTS) -> "GlobalAgent":
         return typing.cast("GlobalAgent", super().with_tts(vendor))
 
+    def with_mllm(self, vendor: BaseMLLM) -> "GlobalAgent":
+        return typing.cast("GlobalAgent", super().with_mllm(vendor))
+
     def with_avatar(self, vendor: BaseAvatar) -> "GlobalAgent":
         return typing.cast("GlobalAgent", super().with_avatar(vendor))