Ax Audio And Realtime For Python

Use when writing Python code with axllm for audio input/output, OpenAI Responses audio mapping, realtime event folding, and generated package audio examples.

Install

Install only this skill for Python:

Shell

npx skills add https://ax-llm.github.io/ax/python/ --skill 'ax-python-audio'

Published skill file: ax-python-audio/SKILL.md.

Source

Source: packages/python/skills/ax-python-audio/SKILL.md
Version: 25.3.2

Skill Instructions

This skill helps an agent write Python code with the generated Ax package axllm. Use the generated package API, examples, and manifests; do not import TypeScript-only APIs unless you are editing the TypeScript package.

When To Use

Map speech, transcription, or realtime events through the generated provider surface.
Use no-key examples for event folding and provider request mapping.
Keep live provider calls behind explicit credentials and provider-api examples.

Package Facts

Language: Python.
Package: axllm.
Package API docs: API.md and axir-api.json.
Capability manifest: axir-capabilities.json.
Runnable examples: examples/.
Real network support: yes.
Scripted no-key transport support: yes.
Runtime profiles: javascript-quickjs, python-pyodide.

Core Pattern

Python

import os
from axllm import ai

llm = ai("openai", api_key=os.environ["OPENAI_API_KEY"])

Speech And Audio Fields

speak() returns TypeScript’s speech result keys: data (base64 audio), format, mimeType, transcript (the spoken text), and sampleRate / channels when the mime type gives them. The older keys audio, mime_type, and sample_rate are removed.
As in TypeScript, a binary speech body’s Content-Type is its mimeType (else the format’s mime type), and a JSON body (by its Content-Type) is read from audio_data, audioData, data, audio.data, output.audio.data, or a Gemini part’s inline data. A JSON body with a bare string under audio, or without an audio data field, raises TypeScript’s error.
Speak requests match TypeScript: OpenAI defaults to gpt-4o-mini-tts with the alloy voice, sends a { id } voice by its id, sends speed when set, and asks for pcm when the format is pcm16; Mistral defaults to voxtral-mini-tts-2603 and sends the voice as voice_id; Grok sends speed when set.
The renderer calls the client’s speak(request, options), which every AxAIService has.
An AxGen audio output field becomes speech with the renderAudio / render_audio option (see the gen skill): the field then holds the speak() result, with the spoken text as its transcript.
An agent’s audio output fields render the same way when renderAudio / render_audio is set on the agent, its responderOptions, or the forward call (the call’s wins); as in TypeScript, the responder’s speak() request takes the forward call’s speech or responderOptions.speech, not a top-level speech option.
An audio input that is a string, or an audio object with a string transcript (a rendered audio output), reaches the model as text. Audio without a transcript goes as an audio part with only its format (wav when it has none) and data.

Relevant API Surface

AxAI: ai, openai_decisions, AxAIOpenAIDecisionsClient, typesafe, AxAITypesafeClient, AxCancellationToken, AxAIServiceAbortedError, get_supported_ai_models, dict[str, str], Callable[[dict[str, str]], dict[str, str]], AIClient.owned_worker_factory, AxChatSession, closable generator, OpenAICompatibleClient, OpenAIResponsesClient, GoogleGeminiClient, AnthropicClient, AxUsageContext, AxUsageEvent, AxUsageObserver, set_usage_observer, AxRuntimeHooks, AxRateLimitInfo, AxRateLimiter, AxTracer, AxMeter, AxGlobals, set_rate_limiter, set_tracer, set_meter, AxBalancer, AxBalancerAdaptiveStrategy, AxBalancerStatsStore, AxInMemoryBalancerStatsStore, create_balancer_route_stats, update_balancer_route_stats, sample_balancer_route_health, MultiServiceRouter, ProviderRouter

Guardrails

Start from package examples for exact native syntax before inventing a new call shape.
Use provider-api examples only when the user explicitly has provider credentials available.
Use no-key examples for deterministic local checks and provider request mapping.
Treat AxIR as the source of generated package truth: if package docs disagree with source code, update the compiler and regenerate packages.
Do not copy repo-maintainer skills from tools/*/skills/ into user packages.