Granola Daily Sync

What It Is

A scheduled Python script that quietly pulls all new Granola meeting notes and full transcripts into a folder of local markdown files. It runs daily via launchd and uses the OAuth tokens the Granola desktop app already stores on my Mac. I built it to test out Claude Cowork and because I wanted my meeting notes to live alongside the rest of my local markdown files, searchable and accessible by any LLM app instead of being locked behind Granola's UI or limited MCP.

What I Learned

Prompting an agent to research before acting changes what it comes back with — My Cowork prompt said "research possible approaches and come back with ideas" before building anything. Opus 4.6 spent its first turn surveying community repos, Granola's API docs, three different MCP implementations, and my own scripts, then came back with three ranked options, including one I hadn't proposed: read the WorkOS tokens Granola already stores locally.
How much an agent can verify its own work depends on what tools you give it — I scoped Cowork to read access on a single folder with no terminal, so when scripts errored on gzip-compressed API responses, duplicate pulls, or nested-list rendering glitches, the agent couldn't see the failures. I had to run the scripts and paste logs back in. With wider local and tool access it probably could have closed the loop without me.
Building something just for yourself is no longer too expensive to justify — One Cowork session plus a few rounds of debugging from me was enough to get this running. Every model release shifts build-vs-buy, and the bar for whether something is worth building at all keeps dropping with it. Truly custom software, built for an audience of one, can finally come off the backlog and ship in an afternoon.
The simplest implementation is often the right one — Now that everything's becoming an agent, it'd be easy to set this up as an agentic workflow instead (such as an OpenClaw cron job or a Claude Code Routine). But the steps never change, so a deterministic Python script is faster, more reliable (see caveat), and doesn't have to re-derive the same plan every day. Caveat: whenever Granola changes their API this handy little script tends to break. I'd like to have an agent empowered to troubleshoot and update itself when that happens.

How It Works

Pulls three things and stitches them into a markdown file per meeting: the notes I typed, Granola's AI-enhanced summary, and the speaker-attributed transcript
Authenticates by reading ~/Library/Application Support/Granola/supabase.json (the same file the Granola desktop app maintains) and refreshes expired WorkOS tokens by extracting the client_id from the JWT
Runs daily at 9pm via macOS launchd; if the Mac is asleep at that hour, it fires on wake
Tracks synced document IDs in .granola_sync_state.json so each run only fetches new meetings. Never overwrites existing files; collisions append _2, _3, etc.

Built With

Claude Cowork running Opus 4.6 did the building, pointed at my prior manual-sync folder (combining scripts originally written with ChatGPT, YAML format, sample outputs) with instructions to research approaches before writing code. Runtime is Python 3 standard library only plus macOS launchd, so /usr/bin/python3 works out of the box with nothing to pip install. ChatGPT Images 2.0 generated hero image.

Loading content...

What It Is

What I Learned

Prompting an agent to research before acting changes what it comes back with — My Cowork prompt said "research possible approaches and come back with ideas" before building anything. Opus 4.6 spent its first turn surveying community repos, Granola's API docs, three different MCP implementations, and my own scripts, then came back with three ranked options, including one I hadn't proposed: read the WorkOS tokens Granola already stores locally.

How much an agent can verify its own work depends on what tools you give it — I scoped Cowork to read access on a single folder with no terminal, so when scripts errored on gzip-compressed API responses, duplicate pulls, or nested-list rendering glitches, the agent couldn't see the failures. I had to run the scripts and paste logs back in. With wider local and tool access it probably could have closed the loop without me.

Building something just for yourself is no longer too expensive to justify — One Cowork session plus a few rounds of debugging from me was enough to get this running. Every model release shifts build-vs-buy, and the bar for whether something is worth building at all keeps dropping with it. Truly custom software, built for an audience of one, can finally come off the backlog and ship in an afternoon.

The simplest implementation is often the right one — Now that everything's becoming an agent, it'd be easy to set this up as an agentic workflow instead (such as an OpenClaw cron job or a Claude Code Routine). But the steps never change, so a deterministic Python script is faster, more reliable (see caveat), and doesn't have to re-derive the same plan every day. Caveat: whenever Granola changes their API this handy little script tends to break. I'd like to have an agent empowered to troubleshoot and update itself when that happens.

How It Works

Pulls three things and stitches them into a markdown file per meeting: the notes I typed, Granola's AI-enhanced summary, and the speaker-attributed transcript

Authenticates by reading ~/Library/Application Support/Granola/supabase.json (the same file the Granola desktop app maintains) and refreshes expired WorkOS tokens by extracting the client_id from the JWT

Runs daily at 9pm via macOS launchd; if the Mac is asleep at that hour, it fires on wake

Tracks synced document IDs in .granola_sync_state.json so each run only fetches new meetings. Never overwrites existing files; collisions append _2, _3, etc.

Built With