Files
2026-07-13 13:39:38 +08:00

374 lines
16 KiB
YAML

# Examples metadata consumed by livekit/agents-jukebox.
#
# Each top-level key under `examples` is the agent name — used as
# - the directory name under examples/
# - the LIVEKIT_AGENT_NAME injected at deploy time
# - the dispatch target in the jukebox's token endpoint
#
# Pixel art icons: `1` = accent color, `2` = lighter accent (highlight
# / secondary detail), `0` = transparent.
version: 1
# All examples deploy to this LiveKit Cloud project. The CI regenerates
# each example's livekit.toml from this subdomain + the per-example
# agent_id below before invoking `lk agent deploy`.
project:
subdomain: examples-wfxyig8v
examples:
frontdesk:
title: Front Desk
description: A receptionist agent wired up to a live calendar. Answers questions about availability, books appointments in real time, and confirms back to the caller.
# brand palette — indigo-500
accent: "#1F44F9"
agent_id: CA_9TqkLsnwhjmE
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/frontdesk"
tags: [tools, scheduling]
icon:
# Calendar with the top header strip in the highlight tone.
size: 12
pixels: |
000000000000
000100010000
001111111000
012222222100
011111111100
010000000100
010000000100
010000000100
010000000100
010000000100
001111111000
000000000000
views:
- rpc: set_appointment_status
title: "\uf073 Schedule"
healthcare:
title: Healthcare
description: A medical front-desk agent that handles intake, authenticates returning patients, books appointments, and hands the conversation off to a human when it needs to.
# LiveKit brand palette (raw-colors.ts) — green-500
accent: "#009E4F"
agent_id: CA_uFGnwowV2PMS
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/healthcare"
tags: [multi-agent, scheduling]
icon:
# Heart with a top-left specular highlight on the left lobe.
size: 12
pixels: |
000000000000
002100011000
022210111100
211111111110
111111111110
111111111110
011111111100
001111111000
000111110000
000011100000
000001000000
000000000000
survey:
title: Survey
description: A structured interview agent that walks candidates through a software engineer screening flow, captures answers as they go, and writes a graded summary to disk.
# brand palette — amber-500
accent: "#EF8B01"
agent_id: CA_APDyCSXwMUNS
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/survey"
tags: [multi-agent]
icon:
# Speech bubble outline with two highlight "text lines" inside.
size: 12
pixels: |
000000000000
001111111100
010000000010
010000000010
010222200010
010000000010
010220000010
010000000010
010111111100
011000000000
010000000000
000000000000
drive-thru:
title: Drive Thru
description: A drive-thru ordering agent that takes the order over voice, builds a structured cart with dynamically-generated tools per menu item, and reads back the total.
# brand palette — red-500
accent: "#FA4C39"
agent_id: CA_cn3jfgKTxXDL
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/drive-thru"
tags: [tools]
icon:
# Side-view car: cabin pillars in accent, windshield + side
# windows in the lighter tone, body and wheels in accent.
size: 12
pixels: |
000000000000
000000000000
001111110000
012221221000
012222111110
100000000001
101110011101
011011110110
001110011100
000000000000
000000000000
000000000000
# The playground reads these `views` and renders a read-only
# markdown card per entry on the right of the orb. The agent fills
# them at runtime by RPC'ing the matching `id` with a markdown
# payload — the card stays hidden until the first payload arrives.
# The agent fills these `views` at runtime by RPC'ing the
# declared method name (e.g. `set_cart_content`) with a markdown
# payload. Each entry renders as a card on the right of the orb,
# hidden until the first non-empty payload, re-hidden on empty.
views:
- rpc: set_cart_content
# Card title is rendered in the body font, which has the
# FA Solid range merged in — so YAML `\u` escapes for any
# FA codepoint render inline. `\uf07a` is shopping-cart.
title: "\uf07a Current order"
inference:
title: Inference
description: A voice agent powered end-to-end by LiveKit Inference. Pick the STT, LLM, and TTS provider live to hear how each one feels in the same conversation.
# brand palette — purple-500
accent: "#9B5CFF"
agent_id: CA_K9e3yQ3RPNKQ
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/inference"
tags: [inference]
icon:
# Three stacked bars: the STT → LLM → TTS pipeline, with the
# middle bar in the highlight tone.
size: 12
pixels: |
000000000000
000000000000
011111111110
011111111110
000000000000
022222222220
022222222220
000000000000
011111111110
011111111110
000000000000
000000000000
# The playground reads these `controls` and renders one widget per
# entry. `type` selects the widget — defaults to `dropdown` for
# backwards-compat with the other examples. Picking an option fires
# an RPC on the agent participant: `rpc` is the method name,
# payload is `{"value": "…"}`. Defaults match the agent's
# DEFAULT_* constants in agent.py.
#
# Supported types:
# - dropdown (default): pill with options list
# - textarea: multiline string editor. Fires `rpc` with the
# current text on every edit-commit (debounced by the FE)
# so the agent can apply it live.
# - link: a button-styled action. On click fires `rpc` on the
# agent (no payload) and opens the returned string as a URL
# in a new tab. The agent owns the URL — useful for deep
# links into Builder, dashboards, etc. where the params
# depend on the agent's current state.
controls:
- rpc: set_system_prompt
type: textarea
label: System prompt
# Keep in sync with INSTRUCTIONS in examples/inference/agent.py.
default: "You're a friendly agent in the LiveKit Playground. The person talking to you is prototyping their own voice agent — they can edit this prompt in the side panel and swap the STT / LLM / TTS models live. Keep replies short, natural, and conversational, and be expressive so they can hear what the selected voice can do. If the conversation lulls or they're not sure what to try, offer to tell them a short joke — and if they say yes, deliver it with good comic timing. If asked which models you're using, answer honestly."
rows: 8
- rpc: set_stt_model
label: STT
default: deepgram/nova-3
options:
- { value: deepgram/nova-3, label: Deepgram Nova-3 }
- { value: deepgram/nova-3-medical, label: Deepgram Nova-3 Medical }
- { value: deepgram/flux-general-en, label: Deepgram Flux (English) }
- { value: deepgram/flux-general-multi, label: Deepgram Flux (Multilingual) }
- { value: assemblyai/u3-rt-pro, label: AssemblyAI Universal-3 Pro }
- { value: assemblyai/universal-streaming, label: AssemblyAI Universal-Streaming }
- { value: assemblyai/universal-streaming-multilingual, label: AssemblyAI Universal-Streaming Multilingual }
- { value: cartesia/ink-whisper, label: Cartesia Ink Whisper }
- { value: elevenlabs/scribe_v2_realtime, label: ElevenLabs Scribe v2 Realtime }
- { value: speechmatics/enhanced, label: Speechmatics Enhanced }
- { value: speechmatics/standard, label: Speechmatics Standard }
- { value: xai/stt-1, label: xAI Speech-to-Text }
- rpc: set_llm_model
label: LLM
default: google/gemma-4-31b-it
options:
- { value: google/gemma-4-31b-it, label: Gemma 4 31B }
- { value: openai/gpt-4.1, label: OpenAI GPT-4.1 }
- { value: openai/gpt-4.1-mini, label: OpenAI GPT-4.1 mini }
- { value: openai/gpt-4.1-nano, label: OpenAI GPT-4.1 nano }
- { value: openai/gpt-4o, label: OpenAI GPT-4o }
- { value: openai/gpt-4o-mini, label: OpenAI GPT-4o mini }
- { value: openai/gpt-5, label: OpenAI GPT-5 }
- { value: openai/gpt-5-mini, label: OpenAI GPT-5 mini }
- { value: openai/gpt-5-nano, label: OpenAI GPT-5 nano }
- { value: openai/gpt-5.1, label: OpenAI GPT-5.1 }
- { value: openai/gpt-5.1-chat-latest, label: OpenAI GPT-5.1 Chat }
- { value: openai/gpt-5.2, label: OpenAI GPT-5.2 }
- { value: openai/gpt-5.2-chat-latest, label: OpenAI GPT-5.2 Chat }
- { value: openai/gpt-5.3-chat-latest, label: OpenAI GPT-5.3 Chat }
- { value: openai/gpt-5.4, label: OpenAI GPT-5.4 }
- { value: openai/gpt-5.4-mini, label: OpenAI GPT-5.4 mini }
- { value: openai/gpt-5.4-nano, label: OpenAI GPT-5.4 nano }
- { value: openai/gpt-5.5, label: OpenAI GPT-5.5 }
- { value: openai/gpt-oss-120b, label: OpenAI GPT OSS 120B }
- { value: google/gemini-2.0-flash, label: Gemini 2.0 Flash }
- { value: google/gemini-2.0-flash-lite, label: Gemini 2.0 Flash-Lite }
- { value: google/gemini-2.5-flash, label: Gemini 2.5 Flash }
- { value: google/gemini-2.5-flash-lite, label: Gemini 2.5 Flash-Lite }
- { value: google/gemini-2.5-pro, label: Gemini 2.5 Pro }
- { value: google/gemini-3-flash-preview, label: Gemini 3 Flash }
- { value: google/gemini-3-pro-preview, label: Gemini 3 Pro }
- { value: google/gemini-3.1-flash-lite, label: Gemini 3.1 Flash Lite }
- { value: google/gemini-3.1-pro-preview, label: Gemini 3.1 Pro }
- { value: xai/grok-4-1-fast-non-reasoning, label: Grok 4.1 Fast }
- { value: xai/grok-4-1-fast-reasoning, label: Grok 4.1 Fast Reasoning }
- { value: xai/grok-4.20-0309-non-reasoning, label: Grok 4.20 }
- { value: xai/grok-4.20-0309-reasoning, label: Grok 4.20 Reasoning }
- { value: xai/grok-4.20-multi-agent-0309, label: Grok 4.20 Multi-Agent }
- { value: deepseek-ai/deepseek-v3, label: DeepSeek-V3 }
- { value: deepseek-ai/deepseek-v3.1, label: DeepSeek-V3.1 }
- { value: deepseek-ai/deepseek-v3.2, label: DeepSeek-V3.2 }
- { value: moonshotai/kimi-k2-instruct, label: Kimi K2 Instruct }
- { value: moonshotai/kimi-k2.5, label: Kimi K2.5 }
- rpc: set_tts_model
label: TTS
default: inworld/inworld-tts-2
options:
- { value: cartesia/sonic, label: Cartesia Sonic }
- { value: cartesia/sonic-2, label: Cartesia Sonic-2 }
- { value: cartesia/sonic-3, label: Cartesia Sonic-3 }
- { value: cartesia/sonic-3-latest, label: Cartesia Sonic-3 Latest }
- { value: cartesia/sonic-turbo, label: Cartesia Sonic Turbo }
- { value: deepgram/aura, label: Deepgram Aura }
- { value: deepgram/aura-2, label: Deepgram Aura-2 }
- { value: elevenlabs/eleven_flash_v2, label: ElevenLabs Flash v2 }
- { value: elevenlabs/eleven_flash_v2_5, label: ElevenLabs Flash v2.5 }
- { value: elevenlabs/eleven_multilingual_v2, label: ElevenLabs Multilingual v2 }
- { value: elevenlabs/eleven_turbo_v2, label: ElevenLabs Turbo v2 }
- { value: elevenlabs/eleven_turbo_v2_5, label: ElevenLabs Turbo v2.5 }
- { value: elevenlabs/eleven_v3, label: ElevenLabs v3 }
- { value: inworld/inworld-tts-1, label: Inworld TTS 1 }
- { value: inworld/inworld-tts-1-max, label: Inworld TTS 1 Max }
- { value: inworld/inworld-tts-1.5-mini, label: Inworld TTS 1.5 Mini }
- { value: inworld/inworld-tts-1.5-max, label: Inworld TTS 1.5 Max }
- { value: inworld/inworld-tts-2, label: Inworld TTS 2 }
- { value: rime/arcana, label: Rime Arcana }
- { value: rime/mist, label: Rime Mist }
- { value: rime/mistv2, label: Rime Mist v2 }
- { value: rime/mistv3, label: Rime Mist v3 }
- { value: xai/tts-1, label: xAI Text-to-Speech }
# "Open in Builder" CTA. `type: link` fires `rpc` on the agent
# and opens the returned string as a URL in a new tab. The
# agent (examples/inference/agent.py :: open_in_builder) reads
# its current instructions + STT/LLM/TTS model strings and
# builds the Cloud Builder deep-link from there, so the
# frontend doesn't have to know the schema.
- rpc: open_in_builder
type: link
label: "Open in Builder \u2192"
primary: true
# Live conversation transcript. The agent doesn't push to this view
# by RPC the way other examples push markdown cards — the frontend
# subscribes to LiveKit `TranscriptionReceived` events and renders
# both sides of the conversation as text. `source: transcript` is
# the trigger; `rpc` is a sentinel value and isn't actually called.
views:
- rpc: __transcript
source: transcript
title: "\uf075 Transcript"
hotel_receptionist:
title: Hotel Receptionist
description: A hotel front-desk agent that books rooms and restaurant reservations, modifies and verifies existing bookings, and confirms guest details — all backed by a live database.
# brand palette — teal-500
accent: "#14B8A6"
agent_id: CA_ZVVNHj3sPjJ5
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/hotel_receptionist"
tags: [multi-agent, tools]
icon:
# Concierge service bell: domed bell with a top-left specular
# highlight, sitting on a full-width base plate.
size: 12
pixels: |
000000000000
000000000000
000001100000
000011110000
000221111000
002211111100
011111111110
011111111110
111111111111
000000000000
000000000000
000000000000
avatar:
title: Avatar
description: Talk to a LemonSlice avatar. Pick anyone from the dropdown to swap who you're chatting with, from Leila to a Pixar fox.
# LemonSlice brand \u2014 yellow-500 from the LK palette.
accent: "#F9E71F"
agent_id: CA_dzjZwsBsRKzZ
entry: agent.py
github: "https://github.com/livekit/agents/tree/main/examples/avatar"
tags: [avatar, video, multi-persona]
# Tells the jukebox frontend to render the AvatarFrame and drive
# the connect-time orb \u2192 video reveal animation. Without this
# flag the example renders like every other one (orb only).
avatar: true
icon:
# Head + shoulders. Head is filled with '2' (highlight tone) and
# the shoulders are '1' (full accent), so the two regions read as
# two distinct shades instead of a single blob.
size: 12
pixels: |
000000000000
000022220000
000222222000
002222222200
002222222200
002222222200
000222222000
000022220000
000000000000
001111111100
011111111110
111111111111
# The playground reads these `controls` and renders one widget
# per entry. The persona dropdown fires `set_avatar` on the
# agent with the chosen id as the payload value. The agent
# updates voice + system prompt live; the avatar image stays
# on the current persona until the session reconnects.
controls:
- rpc: set_avatar
label: Persona
default: leila
options:
- { value: leila, label: Leila }
- { value: jess, label: Jess }
- { value: software_engineer, label: Software Engineer }
- { value: social_worker, label: Social Worker }
- { value: ai_therapist, label: AI Therapist }
- { value: management_consultant, label: Management Consultant }
- { value: shopping_assistant, label: Shopping Assistant }
- { value: cat_girl, label: Cat Girl }
- { value: mr_fox, label: Mr Fox }