When I first published this post four days ago, it covered one MCP server. Then I went on a bit of a spree.
My Codex setup can now search years of podcast transcripts, compare Oura and Garmin data, browse my Plex and Steam libraries, work with my ComfyUI generation queue, and check current OpenAI documentation. Each one is tied to a question I already ask.
Some of the repositories are public and some are staying private. I linked all of them below. The private GitHub links only work for people who already have access.
My current MCP stack
-
Podcast transcript research (private): This was the original project behind this post. I use it to search public captions from Huberman Lab, Cal Newport, Lex Fridman, and Tim Ferriss, then jump to the exact moment on YouTube. It is much faster than guessing an episode and scrubbing through three hours of video.
-
Oura MCP (public): I built a focused, read-only bridge to Oura’s API. I use it to review sleep, readiness, resilience, heart rate, stress, and workouts without handing the agent a general-purpose API client.
-
Garmin MCP (public upstream): Garmin fills in the training side with activities, training readiness, HRV, body battery, stress, and longer-term trends. I run it with a read-only tool allowlist and often compare its view with Oura instead of treating either device as the whole story.
-
Plex MCP (public): This turns my Plex library and watch history into something I can ask questions about. I use it to find media, check what I watched, get recommendations, manage my watchlist and playlists, and record ratings.
-
Steam Library MCP (public): This read-only server covers my complete Steam-returned library, recent playtime, and wishlist. It is useful for questions like “What have I never played?” or “What short game already in my library should I try next?”
-
Comfy Generation Manager MCP (private): This is the control surface for my local ComfyUI system. I use it to inspect workers and failed jobs, search workflows, queue or retry guarded generation jobs, and review outputs while the web app keeps the visual trail.
-
OpenAI Developer Docs MCP (official): I use OpenAI’s read-only documentation server while coding so Codex can search current product and API docs in the same session. It only returns documentation, with no OpenAI API calls on my behalf.
How I use them together
The most useful sessions cross one boundary at a time. I might compare Oura recovery with Garmin training load, use Plex history to choose a movie I have not seen, or inspect a ComfyUI failure and relaunch the right workflow. The agent handles the conversation, while each MCP server owns a narrow source of truth.
I keep personal health data and local services behind stdio or loopback, keep credentials out of prompts and repositories, and default to read-only tools where I can. The private repositories stay private; linking them here does not publish their code or make their MCP endpoints reachable.
That narrowness is what makes the stack useful. Each server has a clear job, and I can tell exactly what data and actions I am giving the agent before I use it.