Podcast providers

The daily-dispatch CLI makes a two-host, call-in style episode from a day's stories in two steps, and each step is an adapter you pick by name:

• A script writer turns the stories into a script: anthropic, openai, ollama or command.
• A voice provider reads the script: elevenlabs, openai, google, polly, azure or local.

ffmpeg then joins the audio into one MP3 next to a Markdown transcript.

Install

npm install -g @blbaker/daily-dispatch
daily-dispatch --help

The package is not on npm yet. Until it is, build it from the repository and run node packages/daily-dispatch/dist/cli/bin.js wherever this page says daily-dispatch:

pnpm install
pnpm --filter @blbaker/daily-dispatch build

It needs Node.js 22.18 or newer, and ffmpeg on your PATH for audio (brew install ffmpeg, apt install ffmpeg or winget install ffmpeg). A dry run works without ffmpeg.

Make an episode

daily-dispatch podcast make --stories stories.json \
  --writer ollama --voice local --voices voices.json --out episode/

episode/ then holds episode.mp3, script.txt, transcript.md and a chunks/ cache. Run it again after editing script.txt with --script episode/script.txt and only the changed chunks are voiced again.

Paid providers never run without --confirm. Without it, the CLI still runs a free writer so you can read the script first, prints the character count and an estimated cost, and sends nothing to a paid service. Add --dry-run to make sure nothing paid runs even with --confirm.

Option	Meaning
--stories <file>	An array of stories, or an edition {"stories": [...]}
--writer <name>	anthropic, command, ollama or openai
--voice <provider>	azure, elevenlabs, google, local, openai or polly
--voices <file>	The voice map, described below
--out <dir>	Output directory
--script <file>	Voice an existing script and skip the writer
--minutes <n>	Target length, default 10
--model, --voice-model	Writer model and voice model
--writer-command <cmd>	Command for --writer command, default claude -p
--style-guide <file>	Replace the built-in style guide

Each story has headline, source, summary and url. The writer is told that every fact must come from those stories.

Script writers

Name	Calls	Key	Default model
anthropic	Messages API	ANTHROPIC_API_KEY	claude-sonnet-5-5
openai	Chat Completions API	OPENAI_API_KEY	gpt-6.1-sol
ollama	Local /api/chat	none	OLLAMA_MODEL, or llama3.1:8b
command	Any CLI, prompt on stdin	none	claude -p

anthropic and openai are paid. command counts as free; what it costs is up to the CLI you point it at. When a model breaks a format rule, the writer sends the problems back and asks for a corrected script.

Voice providers

Name	Service	Key	Voice id example
elevenlabs	ElevenLabs Text-to-Dialogue	ELEVENLABS_API_KEY	an ElevenLabs voice id
openai	OpenAI /v1/audio/speech	OPENAI_API_KEY	ash
google	Gemini multi-speaker TTS	GEMINI_API_KEY	Kore
polly	Amazon Polly, neural by default	AWS credentials and AWS_REGION	Ruth
azure	Azure AI Speech	AZURE_SPEECH_KEY, AZURE_SPEECH_REGION	en-US-AvaMultilingualNeural
local	Any OpenAI-compatible TTS server, such as Kokoro	none	the server's voice name

local talks to http://127.0.0.1:8880/v1 with the model kokoro unless you set LOCAL_TTS_URL and LOCAL_TTS_MODEL. Only ElevenLabs keeps audio tags such as [laughing]; the other providers drop them before sending.

Prices, sources and a per-episode estimate for each provider are on the Daily episode page.

The voice map

voices.json maps each speaker tag in the script to a voice:

{
  "S1": { "id": "voice-id", "name": "Host 1" },
  "S2": { "id": "voice-id", "name": "Host 2" },
  "CALLER_A": { "id": "voice-id", "name": "Caller A" }
}

id is the provider's name for the voice and name appears in the transcript. S1 and S2 are the hosts and any other tag is a caller, unless you set "role". The package ships examples for ElevenLabs and OpenAI in packages/daily-dispatch/examples/. A real run refuses a map that still has the placeholder ids.

Environment variables

API keys are read only from the environment, never from a flag or a config file.

Variable	Used by
ANTHROPIC_API_KEY	--writer anthropic
OPENAI_API_KEY	--writer openai, --voice openai
ELEVENLABS_API_KEY	--voice elevenlabs
GEMINI_API_KEY	--voice google
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, AWS_REGION	--voice polly
AZURE_SPEECH_KEY, AZURE_SPEECH_REGION	--voice azure
LOCAL_TTS_URL, LOCAL_TTS_MODEL, LOCAL_TTS_API_KEY	--voice local
OLLAMA_HOST, OLLAMA_MODEL	--writer ollama, default 127.0.0.1:11434
DAILY_DISPATCH_WRITER_COMMAND	Default for --writer-command
LISTENER_NAME, LISTENER_TOWN	Who the hosts say sent the links
FFMPEG_PATH, FFPROBE_PATH	ffmpeg and ffprobe when they are not on PATH

Add a provider

A voice provider is one file in packages/daily-dispatch/src/voices/ plus one line in that folder's registry.ts. A script writer is one file in src/writers/ plus one line in its registry. The package README has the two interfaces and a worked example of each. Tests pass a fake fetch and check the request, so no test calls a real service.

What has been tested

Every adapter is tested against recorded and faked HTTP responses. The ollama writer and the local voice have also made a full episode end to end with a local model and a local TTS server. No paid service was called while building the CLI, so the request shapes for Anthropic, OpenAI, ElevenLabs, Gemini, Polly and Azure come from their API documentation and have not been tried against the live services.