Published: 2026-09-03

A 5am briefing podcast built from Grok Bot, an MCP voice service and a podcast API

Chapters / key moments (click to jump — plays here on the page)

The transferable technique here is not the podcast — it is the prompt that starts the build. Rather than specifying the whole pipeline up front, Hewitt describes it roughly and ends with "why don't you analyse this and ask me clarifying questions as it relates to this build?" The agent came back with four rounds of questions before writing anything. That is the move worth stealing for any agentic build where you arrive with a half-formed plan. The rest is a genuinely small stack: an agent that can authenticate to an MCP server, a voice service, and a podcast API.

Source video

"How to Build a Custom AI Daily Briefing with Grok Bot, Fish Audio, and Castos" by Craig Hewitt — Watch on YouTube →

Step-by-Step Breakdown

Describe the build roughly, then ask for clarifying questions
The opening prompt describes the goal — a personal briefing pulling calendar, email and anything the chief-of-staff agent has access to, plus industry news — and explicitly asks the agent to analyse it and ask clarifying questions. Four rounds of questions followed, including one the agent was right to ask: is this for the video, or are you actually going to use it? The answer changes the design.
Let the agent find and authenticate the MCP server
The voice service publishes an MCP server; the agent located it and its authentication URL without being told. Authentication runs through the agent's own approval flow — a button in the conversation opens the service dashboard to establish a logged-in session. This is the part that is genuinely new: the credential handshake is a step in the conversation rather than a config file you edit.
Have it create a named, scheduled skill
The agent created a skill called morning podcast and scheduled it for 5am daily. Naming and scheduling in one step is what turns a one-off build into a routine.
Iterate on the voice before you accept the output
The first generated voice was rejected as not worth listening to for ten minutes. Told to design a custom one, the agent logged into the service in a browser as the user, generated two more samples, and presented them as a choice. Reject the first voice; you will hear it every day.
Cut the source that costs money for no benefit
The original plan pulled from X, which meant either developer API credits or a logged-in session. Hewitt dropped it and used general web news sources instead. Worth copying as a habit: when a data source in an agentic build adds cost or an auth dependency without adding much, remove it early rather than engineering around it.
Push to a private feed with a plain API call
The podcast platform had no MCP server, so the agent was pointed at its episode-create API documentation and called it directly. The point made explicitly: for these purposes an API and an MCP server are interchangeable — the absence of an MCP integration is not a blocker. The platform generates a private, individual RSS feed you add to any podcast app.

Gotchas & Caveats

The video is sponsored in part by the voice service, and the creator owns the podcast platform used in the final step. Both are disclosed in the video. The pipeline shape — agent, TTS with an MCP server, a feed with an API — is what transfers; the specific vendors are substitutable.
Giving an agent a browser session that is logged in as you, which is how the custom voice was created, means it acts with your full account privileges on that service. Worth being deliberate about which services you allow that on.
Voice cloning is available on the service used. Clone your own voice freely; do not clone anyone else's.
The briefing draws on calendar and email through a personal agent, so everything in it is as sensitive as your inbox. Treat the generated audio file and its feed URL accordingly — a private RSS URL is a bearer credential.

Key Takeaways

End a half-formed build prompt with "ask me clarifying questions." It makes the agent take half a step back and surface the decisions you had not made — four rounds of them here.
No MCP server is not a blocker. Point the agent at the API docs instead; for a single call the two are equivalent.
The whole build took about twenty minutes, most of it conversation rather than configuration, and produced a roughly five-minute daily briefing.
The briefing's value is context, not bullets — it can say what was last discussed with a given person, that a proposal is outstanding, and that an invoice has gone quiet, because the chief-of-staff agent already holds those threads.
Drop expensive data sources early. The X integration was cut for developer-credit cost and replaced with web news at no loss to the output.
The private-RSS pattern generalises: any agent output you would rather listen to than read can be delivered to a podcast app this way, including to clients.