Codex works from the context you give it. Most people give it their code and stop there. What almost nobody thinks to give it: their meetings.

That's usually because meetings live in some notetaker's cloud, behind a login, owned by a vendor. Weeve breaks that assumption. It records, transcribes and summarizes your meetings entirely on your Mac, and saves everything as ordinary files in a folder on your computer. And ordinary local files are the one thing Codex can use with no plugin, no connector and no setup beyond a note telling it where to look.

Where Weeve keeps your meetings

Open Weeve and look under Settings → Data & Storage. That screen shows (and lets you change) where your meeting library lives. By default it's a folder called Weeve inside your Mac's Application Support directory, with a subfolder per area: Meetings, Recordings, Projects, and so on.

Inside, every meeting is one small file containing four things: the headline, the date, the summary, and the full transcript. Recordings also store who said what and when, line by line, with timestamps.

That simple structure is the whole trick. Codex can skim the short summaries across weeks of meetings, and only open a full transcript when your question really needs the exact words. Fast and cheap by default, precise when it matters.

Tell Codex where to look

Codex reads its standing instructions from a plain text file called AGENTS.md (the spec). Nothing mystical: it's a note you write once, and Codex reads it at the start of every session. If you've used Codex on a project, you may already have one in that project's folder.

Since your meetings aren't tied to one project, put the note in your personal Codex folder, in the file ~/.codex/AGENTS.md, so it applies everywhere. This is the only code-ish step, and it's copy-paste:

## Meeting context (Weeve)

My meeting library (read-only):
~/Library/Application Support/Weeve/Users/<your-user-id>

## Meeting context (Weeve)

My meeting library (read-only):
~/Library/Application Support/Weeve/Users/<your-user-id>

## Meeting context (Weeve)

My meeting library (read-only):
~/Library/Application Support/Weeve/Users/<your-user-id>

The one placeholder is your user ID: it's the name of the single folder you'll find inside Application Support → Weeve → Users. Look it up once in Finder (in the menu bar: Go → Go to Folder → paste ~/Library/Application Support/Weeve/Users), paste it in, done.

What you can ask now

From the next session, Codex answers from your meetings the same way it answers from your code:

• "Read yesterday's meeting summaries and write my standup update."

• "Collect every action item assigned to me from this week's meetings."

• "In Tuesday's recording, what exactly did the customer say about the export feature? Give me the quote."

That last one is the payoff of the line-by-line transcript: you get the real quote with speaker and timestamp, instead of a paraphrase from memory.

The footprint argument

Every cloud notetaker your company adopts grows its digital footprint in three directions at once. A bot appears in your calls, visible to clients and recording them too. A complete archive of your conversations accumulates on a vendor's servers. And your GDPR processor list gets longer. The legal risk is live, not hypothetical: Otter.ai faces federal class-action litigation in California over its bot allegedly recording participants who never signed up (NPR).

The Weeve-plus-Codex setup grows nothing. No bot, because Weeve captures locally without joining the call. No archive anywhere but your Mac. No new vendor, because Codex reading a meeting file sends that content to the same OpenAI service your code already goes to: deliberately, only for the files it opens, under terms you already accepted. Your trust surface stays exactly the size it was.

Two habits keep it clean: leave the read-only rule in place, and never let transcript text land in a repository or commit. Meetings contain other people's words; handle them accordingly.

The result

Your meeting history becomes something Codex can reason over: searchable, quotable, grounded in what was actually said. And the data behind it never left the machine the meeting happened on.

Sources: AGENTS.md specification · NPR: Lawsuit claims AI service Otter secretly records private chats