Thunder Kitty vs Granola

Granola's headline claim is "bot-free AI meeting notes." That's accurate — no bot joins your call. The person on the other end doesn't see a notification. That's a meaningful distinction from tools like Otter or Fireflies.

But bot-free doesn't mean local. When you record a meeting with Granola, your audio goes to Deepgram and AssemblyAI for transcription, then to OpenAI or Anthropic for summaries. Four companies' infrastructure, four terms of service, before you read a word of your notes.

This isn't a scandal — it's just what the product does. Most Granola users are recording team standups and sales calls. The cloud pipeline is fine for that.

But for some meetings, it isn't. A therapy session. A source interview. A coaching conversation. An executive discussion you'd rather not have sitting in someone else's database. Granola wasn't designed for those cases. Thunder Kitty was.

Turn on Airplane mode. Record a meeting. Read the transcript. The app works. That's not marketing — it's a test you can run before you buy anything.

Granola captures system audio as a single channel. On a two-person call, you get a visual left/right split — your bubbles and theirs, unnamed. On a call with three or more people, everyone else gets collapsed into one undifferentiated stream. Granola can't tell them apart because it isn't trying to — it's one audio channel, not separate speakers.

Thunder Kitty does actual speaker diarization. You enroll your own voice once and Thunder Kitty always knows when it's you, labeled correctly in every transcript. Everyone else is separated out — Speaker 1, Speaker 2, Speaker 3 — regardless of how many people are on the call. You can rename them within a transcript ("Speaker 1 → Mike") and that label cascades through that meeting's record.

Thunder Kitty doesn't fingerprint other people's voices across meetings. That would mean collecting biometric data about people who haven't consented to it — and that's not a line we're willing to cross. A small tradeoff for a principle we think matters.

Either way: when you open a Thunder Kitty transcript from a four-person call, you know who said what. In Granola, three of those four voices are one blob.

When a meeting ends in Thunder Kitty, your notes and transcript sit in ~/Documents/Thunder Kitty Notes/ as a plain markdown file. Your handwritten notes and the full timestamped, speaker-attributed transcript, in a format any app can open.

Point Claude at the whole folder and ask questions across every meeting you've ever recorded — no copy-paste, no export, no dashboard. Your meeting archive becomes something you can actually query, in the tools you already use.

Granola's notes live in Granola's database. The company recently changed how it stores data internally — a move that temporarily broke users' AI agent workflows. Their architecture is evolving. Thunder Kitty's architecture is: a folder of text files. There's nothing to break.

Thunder Kitty transcribes using Apple's SpeechAnalyzer — the same on-device engine powering Apple Intelligence, running entirely on your Mac. No cloud billing. No per-minute meter.

Record for 6 hours. Record 12 meetings in a day. Record yourself thinking through a problem out loud. No cap, no warnings, because the cost to us is zero.

Granola's free plan limits meeting history to 30 days. Unlimited history — and access to better AI models — requires the $14/month Business plan.

Granola uses GPT-4 and advanced thinking models for summaries. These are genuinely powerful and we won't pretend otherwise.

Thunder Kitty ships with the best available on-device models, updated as new releases come out. On-device models are improving fast — for structured use cases like meeting summaries, the gap with cloud models narrows every few months. If you want even stronger summaries, plug in your own Claude, OpenAI, or Gemini API key. You pay the provider directly for exactly what you use, with your own key, under your own terms.

For transcription: Apple's SpeechAnalyzer on macOS 26 is excellent. Fast, accurate, and it never sends a byte of audio anywhere.