Daily podcast episode

Daily Dispatch can make a podcast episode from each day's stories at the same time every day. It is off by default and nothing runs until you set it up.

The setup has two parts that stay separate on purpose:

• Settings say whether the episode is on, what time it is made, in which time zone, and which script writer, voice provider and host voices to use. They live in the app's Settings screen (stored by the backend at /settings) or, without a backend, in a local config file.
• A scheduled job runs daily-dispatch podcast run --if-due every 15 minutes. Each run reads the settings, and does nothing unless the episode is on, the local time is past the set time, and today has no episode yet.

Because the job reads the settings on every run, changing the time in the app takes effect at the next check. You never reinstall the job to change the time.

Turn it on

Install the CLI and ffmpeg on the machine that will make the episodes:

npm install -g @blbaker/daily-dispatch
ffmpeg -version            # macOS: brew install ffmpeg, Debian/Ubuntu: apt install ffmpeg

Install it globally rather than running it through npx. The scheduled job stores the path of the CLI script, and the npx cache can be cleared.

Where the settings come from

If you run the backend, set these two variables and the job uses whatever the app's Settings screen saved:

DAILY_DISPATCH_API_URL=https://<your function URL>
DAILY_DISPATCH_API_KEY=<the backend API key>

These are the same values the pipeline calls DIGEST_API_URL and DIGEST_API_KEY. The day's stories then come from the backend's /editions/<date>.

Without a backend, put the settings in ~/.config/daily-dispatch/config.json ($XDG_CONFIG_HOME/daily-dispatch/config.json if that is set, or any file you pass with --config):

{
  "podcast": {
    "enabled": true,
    "time": "06:30",
    "timezone": "America/Chicago",
    "writer": "ollama",
    "voice": "local",
    "voices": { "S1": { "id": "af_bella", "name": "Bella" }, "S2": { "id": "am_adam", "name": "Adam" } }
  },
  "stories": "~/digests/{date}.json",
  "out": "~/Podcasts/daily-dispatch",
  "minutes": 10,
  "confirm": false
}

Field	Meaning
podcast.enabled	The on/off switch. Default false.
podcast.time	Local time, 24-hour HH:MM. Default 06:00.
podcast.timezone	IANA zone such as America/Chicago. An unknown zone is an error; it is never treated as UTC.
podcast.writer	command (default, runs claude -p), ollama, anthropic or openai.
podcast.voice	elevenlabs (default), openai, google, polly, azure or local.
podcast.voices	A voice map like voices.json, a path to one, or host name to voice id ({"Bella": "af_bella"}, which becomes S1, S2 in order).
stories	Local digest file; {date} becomes YYYY-MM-DD. Leave it out to read the backend's edition.
out	Episodes go in <out>/<date>/. Default ~/.local/share/daily-dispatch/episodes.
minutes, title, listener	Episode length, show title, and who sent the links.
model, voiceModel, writerCommand	The same as --model, --voice-model and --writer-command on podcast make.
confirm	Must be true before a scheduled run calls a paid provider. Default false.

When the backend is configured, its settings decide enabled, time, timezone, the writer, the voice provider and the host voices. The config file still supplies the local fields (stories, out, confirm and so on), and its podcast.voices fills in when the app has no host voices saved.

API keys

Schedulers do not read your shell profile, so keys set in ~/.bashrc are not visible to the job. Put them in ~/.config/daily-dispatch/env instead, one KEY=VALUE per line, and make the file readable only by you:

chmod 600 ~/.config/daily-dispatch/env

A variable that is already set in the environment wins over the file. The variables each provider reads are listed in the package README.

Try it by hand first

daily-dispatch podcast run --dry-run     # prints characters and cost, calls nothing paid
daily-dispatch podcast run               # makes today's episode now
daily-dispatch podcast run --if-due      # what the job runs

podcast run without --if-due ignores the on/off switch and the time, but still makes at most one episode per date unless you add --force.

Install the job

daily-dispatch schedule install --print  # shows the files and commands, changes nothing
daily-dispatch schedule install          # registers the job
daily-dispatch schedule status

--every 30m changes how often the job checks. The default is 15 minutes.

The schedulers

schedule install picks the scheduler for the machine. Pass --kind to choose one yourself, and pass the same --kind to uninstall and status.

Every job runs node <path to the CLI> podcast run --if-due with the PATH you had when you installed it, so it can find ffmpeg and any writer CLI. If you move ffmpeg or reinstall Node somewhere else, run schedule install again. All of them append to ~/.local/state/daily-dispatch/podcast.log.

systemd (Linux, preferred when systemd is running)

Writes ~/.config/systemd/user/daily-dispatch-podcast.service and daily-dispatch-podcast.timer, then runs systemctl --user daemon-reload and systemctl --user enable --now daily-dispatch-podcast.timer. The timer uses OnCalendar with Persistent=true, so a laptop that slept through a check runs it on wake.

User timers stop when you log out unless lingering is on:

loginctl enable-linger "$USER"
systemctl --user list-timers daily-dispatch-podcast.timer

cron (Linux without systemd, or --kind cron)

Adds one line to your crontab between two marker comments. Reinstalling replaces those lines and uninstalling removes only them. cron only repeats at intervals that divide an hour or a day evenly (5m, 10m, 15m, 20m, 30m, 1h, 2h and so on). cron can start a new check while a long run is still going; a lock file next to the state file makes the second one skip.

launchd (macOS)

Writes ~/Library/LaunchAgents/dev.daily-dispatch.podcast.plist with StartInterval set to the interval, and loads it with launchctl bootstrap gui/<uid>. It runs while you are logged in.

Task Scheduler (Windows)

Writes daily-dispatch-podcast.xml next to the config file and registers it with schtasks /Create /TN DailyDispatch\Podcast /XML <file> /F. The task repeats every interval, starts late if the computer was off, and never runs two copies at once. It runs while you are logged in. Task Scheduler cannot redirect output, so the task passes --log and the CLI writes the log itself.

GitHub Actions (no always-on machine)

Copy templates/github-actions-daily-podcast.yml to .github/workflows/daily-podcast.yml in a repository you own, commit a config file at daily-dispatch/config.json, and add your keys as repository secrets (AWS_REGION and AZURE_SPEECH_REGION as variables). The workflow checks once an hour, keeps the state file in the Actions cache, and uploads each episode as an artifact for 14 days. With the DAILY_DISPATCH_API_URL and DAILY_DISPATCH_API_KEY secrets set, it also publishes the episode to the app.

Hosted runners have no Ollama, no local TTS server and no claude CLI, so use the anthropic or openai writer and a cloud voice, with "confirm": true. GitHub can start scheduled runs late and pauses them after 60 days without a commit on a public repository. Only the state file is cached, so a run that fails partway writes a new script on its next try.

How the time in Settings takes effect

Each check reads the settings fresh, so:

• Turning the episode off in the app stops it from the next check. The job keeps running and logs not due: disabled.
• Changing the time or time zone applies to the next check. If the new time has already passed today and today has no episode, the next check makes one.
• The episode belongs to the local date in the settings' time zone, which can differ from the UTC date and from the machine's local time zone.
• A time inside the hour that daylight saving skips (02:30 on a spring-forward day in the US) runs at the first check after the clock jumps. The hour that repeats in autumn runs once, because the date is recorded.
• If the machine was asleep or off at the set time, the next check that same local day makes the episode. A day that passes completely is not made later.

Publishing the episode to the app

When DAILY_DISPATCH_API_URL and DAILY_DISPATCH_API_KEY are set, a run that makes an episode also publishes it to the backend:

1. POST /episodes with the date as the id, a title such as Link Talk: Monday, October 19, 2026, and the length in seconds. The backend replies with a presigned upload URL.
2. PUT of episode.mp3 to that URL. The API key is not sent with it.
3. POST /episodes with {"id": "<date>", "ready": true}. The backend checks that the audio arrived, marks the episode ready and sends the "New episode ready" push.

The app finds episodes through the backend. GET /episodes lists the ready ones, newest first, and GET /episodes/<date> returns one; both include an audio URL that works for an hour. The push carries data.type episode, data.id (the date) and data.path (/episodes/<date>), so a tap can open that episode. Without the two variables, and on a --dry-run, nothing is uploaded. Episodes made while the variables were unset are not sent later.

When something goes wrong

The state file, ~/.local/state/daily-dispatch/state.json, records each date as done or failed with an attempt count, plus the upload's progress when the backend is set.

• No digest for the date yet, a 404 from /editions/<date>, or an edition still in progress means "waiting". The next check tries again and no attempt is counted.
• A failure after spending could start (a provider error, a network error) counts an attempt. A retry reuses the script from the earlier attempt and the chunk audio already made, so you do not pay for those twice. After three failed attempts the job stops trying until the next date.
• A configuration problem (missing key, unknown provider, no ffmpeg, a paid provider without "confirm": true) fails before anything is spent and does not count as an attempt. It shows in the log on every check until fixed.
• A failed upload does not count as an attempt either, because the audio exists. The run exits 1, the upload stays pending, and every later check retries only the upload until it works. If episode.mp3 is gone by then (a GitHub Actions runner keeps only the state file), the upload is marked gave-up, and podcast run --force --date <date> makes and sends it again.
• daily-dispatch podcast run --force makes today's episode regardless of the state file.

daily-dispatch schedule status shows whether the job is installed, the log path, and the last three dates.

Turn it off

To pause it, switch the daily podcast off in the app, or set "enabled": false in the config file. The job keeps checking and does nothing.

To remove the job:

daily-dispatch schedule uninstall        # add --kind if you installed with one

Episodes, the config file, the state file and the log stay where they are. For the GitHub Actions workflow, delete the workflow file or disable it in the repository's Actions tab.

Costs

Paid providers are only called when "confirm": true is set, and podcast run --dry-run prints the estimate for your stories without calling them. The per-episode figures below assume a 10-minute episode of about 10,000 characters (the CLI estimates 1,000 characters per minute of speech). Prices are list prices checked on 2026-10-19; free tiers and plan credits lower them.

Voice provider	List price	10-minute episode	Source
elevenlabs (Eleven v4, Text-to-Dialogue)	$0.08 per 1,000 characters	about $0.80	elevenlabs.io/pricing/api
openai (gpt-4o-mini-tts)	about $0.015 per minute of audio	about $0.15	developers.openai.com/api/docs/pricing
google (gemini-3.8-flash-tts)	$9.00 per million audio tokens, $0.00225 per 10 s of audio; doubles on 2027-01-01	about $0.14	ai.google.dev/gemini-api/docs/pricing
polly neural (default engine)	$16 per million characters	about $0.16	aws.amazon.com/polly/pricing
polly generative / long-form / standard	$30 / $100 / $4 per million characters	about $0.30 / $1.00 / $0.04	aws.amazon.com/polly/pricing
azure (neural voices)	$15 per million characters; the free F0 tier includes 0.5 million a month	about $0.15	azure.microsoft.com/pricing/details/speech
local (Kokoro or any OpenAI-compatible server)	free	$0.00	runs on your machine

Script writer	List price	Per episode	Source
anthropic (claude-sonnet-5-5)	$2 input / $10 output per million tokens	a few cents, printed by --dry-run	platform.claude.com models overview
openai (gpt-6.1-sol)	$2 input / $10 output per million tokens	a few cents, printed by --dry-run	developers.openai.com/api/docs/pricing
ollama	free	$0.00	runs on your machine
command (claude -p by default)	whatever that CLI costs you	not estimated	your CLI's plan

The GitHub Actions workflow is free on a public repository. On a private repository it uses about a minute per hourly check, roughly 750 of the 2,000 minutes a month included with GitHub Free (docs.github.com, "GitHub Actions billing").

The Gemini figure works out as 1.5 audio tokens per character at 1,000 characters per minute, plus a quarter of a text token per character at $0.50 per million. The Polly and Azure figures are characters times the list price. None of these numbers come from a real billed run; no paid API was called while building this.