Pipecat Anam Integration

Generate real-time video avatars for your Pipecat AI agents with Anam.

Maintainer: Anam (@anam-org)

Installation

pip install pipecat-anam

Or with uv:

uv add pipecat-anam

You'll also need Pipecat with the services you use (STT, TTS, LLM, transport). For this repo's examples:

uv sync --extra dev --extra example

That installs all required Pipecat extras (deepgram, cartesia, google, daily, runner, webrtc) plus local tooling.

If you prefer pip:

pip install -e ".[dev,example]"

If you are building your own pipeline, install only the Pipecat extras you need.

Prerequisites

• Anam API key
• API keys for STT, TTS, and LLM (e.g., Deepgram, Cartesia, Google)
• A Daily.co room and (optional) meeting tokens for Daily WebRTC transport — see Auto-provisioning the Daily room.

Usage with Pipecat Pipeline

The AnamVideoService wraps around Anam's Python SDK for a seamless integration with Pipecat to create conversational AI applications where an Anam avatar provides synchronized video and audio output while your application handles the conversation logic. The AnamVideoService iterates over the (decoded) audio and video frames from Anam and passes them to the next service in the pipeline.

enable_audio_passthrough=True renders the avatar directly from your TTS audio (no separate Anam-side LLM or voice generation).

enable_session_replay=False disables Anam-side session recording.

from anam import PersonaConfig
from pipecat_anam import AnamVideoService

persona_config = PersonaConfig(
    avatar_id="your-avatar-id",
    avatar_model="cara-4",
    enable_audio_passthrough=True,
)

anam = AnamVideoService(
    api_key=os.environ["ANAM_API_KEY"],
    persona_config=persona_config,
    api_base_url="https://api.anam.ai",
    api_version="v1",
)

pipeline = Pipeline([
    transport.input(),
    stt,
    context_aggregator.user(),
    llm,
    tts,
    anam,  # Video avatar (returns synchronized audio/video)
    transport.output(),
    context_aggregator.assistant(),
])

See examples/video-avatar-anam-video-service.py for a complete working example.

Director notes

Director notes guide how the avatar performs its speech — a baseline delivery style and expressivity for the session. Configure them on PersonaConfig(director_notes=...) to set the default performance for the session.

from anam import DirectorNotes, PersonaConfig

persona_config = PersonaConfig(
    avatar_id="your-avatar-id",
    avatar_model="cara-4",
    director_notes=DirectorNotes(
        preset_style="warm",
        expressivity=0.7,
    ),
    enable_audio_passthrough=True,
)

Director notes require a Cara-4 avatar. For the supported preset styles, expressivity range, and runtime cues, see the Anam Director Notes docs.

Sending cues during the conversation

To shift the delivery for the current turn, send a cue while the session is active using send_director_note_cue(...):

await anam.send_director_note_cue("laughter", at_seconds=0.0)

The cue applies to the current turn only: use at_seconds for an offset from the start of the turn, or in_seconds for a delay from now.

Avatar model

Set avatar_model on PersonaConfig (default: "cara-4"). If your avatar does not support Cara-4, set avatar_model="cara-3", or create a new compatible avatar in lab.anam.ai.

Output video dimensions (portrait / landscape)

Both AnamVideoService and AnamTransport accept optional video_width and video_height to control the avatar's output resolution. Both must be provided together.

Orientation	Model	Width	Height
Portrait	cara-4	768	1152
Landscape	cara-4	1152	768
Landscape	cara-3	720	480

transport = AnamTransport(
    api_key=os.environ["ANAM_API_KEY"],
    persona_config=persona_config,
    daily_room_url=os.environ["DAILY_ROOM_URL"],
    video_width=768,
    video_height=1152,  # portrait
)

examples/video-avatar-anam-transport.py defaults to landscape via ANAM_VIDEO_WIDTH / ANAM_VIDEO_HEIGHT (1152×768). Set 768 and 1152 for portrait.

AI avatar disclosure watermark

Use this when you want to disclose to the user that they're talking to an AI avatar via a watermark. Both AnamVideoService and AnamTransport accept an optional show_ai_avatar_disclosure bool, forwarded as-is to the Anam SDK's session options. Anam default is False.

transport = AnamTransport(
    api_key=os.environ["ANAM_API_KEY"],
    persona_config=persona_config,
    daily_room_url=os.environ["DAILY_ROOM_URL"],
    show_ai_avatar_disclosure=True,
)

Session region

Both AnamVideoService and AnamTransport accept optional region and region_policy session options (Enterprise plans). See Session Regions for available regions, policies, and plan requirements — additional regions may be introduced over time.

transport = AnamTransport(
    api_key=os.environ["ANAM_API_KEY"],
    persona_config=persona_config,
    daily_room_url=os.environ["DAILY_ROOM_URL"],
    region="eu",  # example; see docs for available regions
    region_policy="strict",
)

region on its own is a preference: the default region_policy="preferred" lets the session fall back to another region when the requested one has no capacity, and the fallback is logged as a warning. Use region_policy="strict" when you need a data-residency guarantee — the connection then fails rather than falling back. region_policy="strict" without a region raises ValueError.

In the AnamTransport path only the Anam session is pinned — the Daily room is yours, so its own media routing is unaffected. Pin the Daily room separately if the whole call has to stay in one region.

Initializing the Anam avatar session

AnamVideoService opens its connection to the Anam Backend asynchronously. The StartFrame is propagated downstream immediately so the rest of the pipeline (LLM/TTS/...) can warm up in parallel. TTS audio starts forwarding once the avatar is ready; any TTS produced before then is held back so it doesn't get dropped on the way in or accumulates latency.

Prior to v0.0.4, AnamVideoService blocked on StartFrame until the avatar was ready, which serialised pipeline startup. The async path keeps initial response latency low.

Publishing directly to Daily

Warning

Direct Daily egress is experimental and only supported for Cara-4 avatars. The transport and signalling path will change in upcoming anam alpha releases. Pin to an exact alpha if you build on this; expect breaking changes between alphas.

AnamTransport is a drop-in replacement for Pipecat's DailyTransport. The Anam Backend publishes the avatar's synchronised audio and video directly into your Daily room, so the Pipecat bot doesn't have to receive and re-publish the avatar's A/V tracks.

The Daily room is bring-your-own: provision the room and mint two separate meeting tokens before starting the pipeline. See the Daily REST API docs for rooms and meeting-tokens (or use pipecat's Daily helpers).

• daily_avatar_token — for the Anam Backend (optional, but required for private rooms). If a user_name claim is set, it must match daily_avatar_user_name (or leave the claim empty). This lets the transport tell the avatar apart from end users. The transport will not forward TTS until the avatar has joined.
• daily_bot_token — for the Pipecat bot itself, used to capture the user's microphone for STT.

Requires anam==0.8.0 (pinned exactly — see the SDK's experimental-alpha warning).

from anam import PersonaConfig
from pipecat_anam import AnamTransport

transport = AnamTransport(
    api_key=os.environ["ANAM_API_KEY"],
    persona_config=PersonaConfig(
        avatar_id=os.environ["ANAM_AVATAR_ID"],
        # Direct Daily egress requires a cara-4 avatar; stock avatars default to cara-3.
        avatar_model="cara-4",
        enable_audio_passthrough=True,
    ),
    daily_room_url=os.environ["DAILY_ROOM_URL"],
    daily_bot_token=os.environ["DAILY_BOT_TOKEN"],
    daily_avatar_token=os.environ["DAILY_AVATAR_TOKEN"],
    daily_avatar_user_name=os.environ["DAILY_AVATAR_USER_NAME"],
)

Auto-provisioning the Daily room

AnamTransport does not mint Daily rooms or tokens itself. If you'd rather provision a room programmatically than pre-create one, use Pipecat's DailyRESTHelper with your DAILY_API_KEY to create the room and the two meeting tokens before constructing the transport:

import aiohttp
from pipecat.transports.daily.utils import (
    DailyMeetingTokenParams,
    DailyMeetingTokenProperties,
    DailyRESTHelper,
    DailyRoomParams,
)

async with aiohttp.ClientSession() as session:
    helper = DailyRESTHelper(
        daily_api_key=os.environ["DAILY_API_KEY"],
        aiohttp_session=session,
    )
    room = await helper.create_room(DailyRoomParams())
    avatar_token = await helper.get_token(
        room.url,
        params=DailyMeetingTokenParams(
            properties=DailyMeetingTokenProperties(user_name="anam-avatar"),
        ),
    )
    bot_token = await helper.get_token(room.url)

    transport = AnamTransport(
        api_key=os.environ["ANAM_API_KEY"],
        persona_config=PersonaConfig(
            avatar_id=os.environ["ANAM_AVATAR_ID"],
            # Direct Daily egress requires a cara-4 avatar; stock avatars default to cara-3.
            avatar_model="cara-4",
            enable_audio_passthrough=True,
        ),
        daily_room_url=room.url,
        daily_avatar_token=avatar_token,
        daily_bot_token=bot_token,
    )

Using the Pipecat runner (recommended for Pipecat Cloud)

When the Pipecat runner dispatches the session (locally via pipecat.runner.run.main, or in Pipecat Cloud), it hands the bot a DailyRunnerArguments with the Daily room URL and a single meeting token for the bot itself. AnamTransport needs two tokens (one for the bot, one for the avatar), so the bot mints the second one in-process via DailyRESTHelper keyed on DAILY_API_KEY. The same file runs unchanged both locally and on Pipecat Cloud.

See examples/video-avatar-anam-transport-pcc.py for the full example.

Video Post-Filter Example

The output transport scales the avatar resolution to the configured output resolution. When the aspect ratios mismatch the video is stretched or squeezed. To avoid this, apply a video post-processing filter that crops the avatar to the output aspect ratio.

examples/video-avatar-anam-postfilter.py adds a CenterAspectCropFilter after AnamVideoService:

• Works on OutputImageRawFrame; does not depend on Anam internals.
• Assumes packed RGB24 bytes (format="RGB").
• Performs a centered crop to match the configured output aspect ratio.
• Does not scale. Pipecat's output transport can still scale as needed.
• No-op when source and target aspect ratios already match.

The filter is self-contained in that file and can be lifted into any Pipecat pipeline that produces OutputImageRawFrame.

Running the Example

1. Install dependencies:

uv sync --extra dev --extra example

2. Set up your environment:

cp env.example .env
# Edit .env with your API keys

3. Run the AnamVideoService example (Pipecat's built-in transports):

uv run python examples/video-avatar-anam-video-service.py -t daily

Or with the built-in WebRTC transport:

uv run python examples/video-avatar-anam-video-service.py -t webrtc

To run the AnamTransport example (direct Daily egress, BYO room and tokens, Deepgram + Google + Cartesia):

uv run python examples/video-avatar-anam-transport.py

Defaults to landscape (1152×768); set ANAM_VIDEO_WIDTH=768 and ANAM_VIDEO_HEIGHT=1152 for portrait.

To run the Pipecat-Cloud-shaped AnamTransport example (same pipeline, but the room is minted by the runner and the avatar token is minted in-process). Use -d so the runner prints a ready-to-click Daily URL:

uv run python examples/video-avatar-anam-transport-pcc.py -d

To run the center-aspect post-filter example with the WebRTC transport:

uv run python examples/video-avatar-anam-postfilter.py -t webrtc

Or with the Daily transport:

uv run python examples/video-avatar-anam-postfilter.py -t daily

Compatibility

• Tested with Pipecat v0.0.100+
• Python 3.10+
• Daily transport or built-in WebRTC transport

License

BSD-2-Clause - see LICENSE

Support

• Anam Lab (Build and test your persona and get your avatar_id.)
• Anam Documentation (API reference and SDK documentation)
• Anam Community Slack
• Pipecat Discord (#community-integrations)