The real fight over AI note-takers isn’t accuracy — it’s custody of your archive
Every social media operator I know is sitting on a graveyard of half-processed audio. Interview clips, voice memos from a shoot, a TikTok voiceover you recorded in one take and never transcribed, the 40-minute call where a brand finally agreed to your rate. The recording tools we use decide what happens to that material — and right now, most of them quietly decide that it lives on their servers, gets processed by their model of choice, and stays behind their export button. SaidTrace, launching on Product Hunt from maker Tiger, is a small but pointed argument against that default: record on your device, pick your own AI, keep the archive on your own Mac. Whether or not you install it, the design tradeoffs it exposes are the ones every creator tooling decision now hinges on.
What SaidTrace actually is, minus the launch-page gloss
Stripped of positioning, SaidTrace is a Mac-based recording archive with three moving parts. First, ingestion: you bring audio in from an iPhone, an Apple Watch, a standalone recorder, or plain audio files, and it lands in one library. Second, processing: it turns recordings into transcripts and notes using either a local LLM via LM Studio or Ollama, or an external AI of your choosing. Third, retrieval: you search the archive, ask questions of it, and — this is the part most tools fudge — follow any answer back to the specific recording and transcript it came from.
There’s also an MCP connector, so external tools like Codex and Cursor can search your notes and pull transcripts. The storage claim is the headline: the team says SaidTrace doesn’t store your audio, transcripts, or notes on its servers, Apple recordings can sync through your own iCloud, and any external AI only receives the text needed for a given request. Pricing is free to start with 20 imports and 20 AI summaries per month, with planned paid tiers at $3/month or $18/year — hardware and external AI fees billed separately.
That’s the whole product. It is deliberately unglamorous, and that’s the point.
Why “bring your own model” is a bigger deal than it sounds
Most AI note-takers bundle the model into the subscription. You pay Otter or Fireflies or Granola, and the transcription engine is whatever they’ve chosen, updated on their schedule, priced into their margin. That’s fine until it isn’t: when a model gets swapped and your summaries change character, when a vendor’s privacy policy shifts, or when you want to run a sensitive client call through something that never leaves your laptop.
SaidTrace inverts that. The model is a slot you fill. In my experience testing local-LLM workflows, that flexibility is genuinely useful for anyone handling embargoed brand material, unreleased product footage, or talent interviews where the subject hasn’t consented to cloud processing. It’s also, frankly, a support burden — more on that below.
How it stacks up against the tools you’re probably already paying for
If you’re a social media manager, your audio probably already flows through one of four buckets, and it’s worth being honest about which.
Meeting-first tools. Otter, Fireflies, and Fathom are built around calendar integrations and shared team workspaces. They’re excellent at the thing SaidTrace doesn’t attempt: joining a Zoom automatically, tagging speakers, pushing action items into Slack. If your job is mostly internal syncs, none of this is for you.
Voice-memo-first tools. Granola and AudioPen target the “I think out loud and want it cleaned up” workflow. Lighter, faster, cloud-dependent.
Creator-native tools. Descript is the real incumbent here — it’s not just transcription, it’s editing by transcript, which is how a lot of short-form video gets cut now. CapCut handles auto-captions inside the edit. If your audio’s destination is a Reel or a Short, these are the tools that actually ship content.
Local/private tools. MacWhisper proved there’s a real market for on-device transcription on Mac. Obsidian plus a local model is the DIY version of what SaidTrace is productizing.
My take: SaidTrace’s closest philosophical cousin is MacWhisper, and its closest functional competitor is whatever combination of a recorder app, a transcription tool, and a notes app you’ve duct-taped together. The pitch is that the duct tape is now one product with a searchable, citation-backed archive and an MCP door for other AI tools.
Where the “your archive, your AI” pitch actually bites
Here’s the operational scenario where this matters, and it’s not hypothetical for anyone running creator partnerships. You record a 30-minute call with a brand about a sponsored series. Six weeks later, legal asks what was promised. Six months later, you want to pull the exact quote where they approved a specific claim for an on-screen caption.
In a cloud tool, that retrieval depends on the vendor still existing, still offering that tier, and still letting you export. In a local archive with citations back to the source recording, it depends on your hard drive. For solo creators and small teams, that’s a meaningfully different risk profile. I’d bet the “follow the answer back to the recording” feature — not the local LLM — is what actually converts people, because citation is the thing every AI summary tool is worst at.
What social media teams can steal from this, regardless of whether they switch
You don’t need to migrate your stack to learn something from how SaidTrace is built. Three transferable ideas:
1. Separate capture from processing from storage. Most creators conflate all three because the tools bundle them. The moment you treat them as separate decisions — where audio lands, which model summarizes it, where the archive lives — you stop being locked in. Even if you stay on Notion plus Rev, naming the three layers makes vendor swaps survivable.
2. Insist on citations in every AI output. If a tool summarizes your content and can’t show you the source, it’s generating, not retrieving. For repurposing workflows — turning one long interview into ten short-form hooks — that distinction is the difference between a usable draft and a fabrication you have to fact-check from scratch.
3. Treat your raw archive as an asset, not exhaust. The recordings you made this year are the raw material for next year’s content. A searchable, locally-held archive with an MCP endpoint means your existing AI tools — Claude, ChatGPT, whatever you use for drafting — can query your own back catalog. That’s a repurposing engine most creators haven’t built because their audio is scattered across five apps.
Why TikTok and YouTube creators should care more than LinkedIn ones
If your output is written — LinkedIn posts, newsletters, X threads — your raw material is already text, and transcription is a nice-to-have. If your output is video or audio, transcription is your production pipeline. Captions, hooks, chapter markers, SEO descriptions, clip selection: all downstream of a transcript. A local, searchable, citable transcript archive is more valuable to a TikTok or YouTube creator than to a B2B writer, full stop.
The counterargument: those creators also need the transcript to flow into an editor, and SaidTrace is an archive, not an editor. It hands you text; it doesn’t cut video.
Where I think SaidTrace falls short, and who shouldn’t bother
Launch-day pages are optimistic by nature, so let me be the wet blanket.
Mac-only, and that’s a real constraint. The archive lives on your Mac. If your team is on Windows, or you edit on an iPad, this is a non-starter today. The maker doesn’t disclose a Windows or mobile roadmap.
“Local LLM” is a hobbyist’s phrase, not a consumer’s. Running Ollama or LM Studio requires comfort with model downloads, quantization, and hardware limits. Anyone who’s done it knows it’s fifteen minutes of setup and occasional debugging. Anyone who hasn’t will bounce. The product’s flexibility is also its onboarding cliff.
The free tier is a taste, not a workflow. 20 imports and 20 AI summaries per month is roughly one busy week for an active creator. The paid tier — $3/month or $18/year — is cheap enough that it’s almost a non-decision, but note that hardware and external AI costs are separate, so the true cost depends on whether you’re running local (electricity and a decent machine) or paying per-token to an external provider.
No disclosed sync story beyond iCloud, and no disclosed team features. The maker says Apple recordings can sync through your own iCloud. Whether that means multi-device access, shared archives, or anything resembling team collaboration is not disclosed. For solo operators, fine. For agencies, unclear.
The MCP angle is early. MCP connectors are powerful and also still settling as a standard. Betting a workflow on one is reasonable in 2025; betting a business on one is not.
Who it’s not for: teams that live in shared cloud workspaces, anyone who needs automatic meeting capture, anyone on Windows, and anyone who wants the tool to also edit video. That’s a lot of people. The addressable audience here is narrower than the launch page implies.
What I’d watch, and what I’d test this week
The interesting question isn’t whether SaidTrace beats Otter. It’s whether “local archive plus bring-your-own-model plus MCP” becomes the default architecture for creator tooling, or stays a niche for privacy-minded power users. My bet is the former, slowly — because the same argument is being made across the stack, from Obsidian to local image generation, and because platform API terms keep getting more restrictive, which pushes creators toward owning their own data.
Concretely, here’s what I’d do this week, whether or not you install it:
- Audit where your audio actually lives. List every app that holds a recording you made in the last 90 days. If the answer is more than three, you have an archive problem before you have a tool problem.
- Test one local transcription run. Install MacWhisper or stand up Ollama and transcribe a single 20-minute recording. You’ll learn in an hour whether local is viable for your hardware and patience.
- Try the citation test on your current tool. Ask it a specific question about something said in a recording and see if it can point you to the timestamp. If it can’t, that’s the gap SaidTrace is selling into.
- If you’re on Mac and privacy-sensitive, take the free tier for a spin and see whether the MCP connector plays nicely with the AI tools you already use. The $18/year tier is cheap enough to be an experiment, not a commitment.
The recording app you use shouldn’t decide where your data lives or which AI you’re allowed to think with. That’s the maker’s framing, and I think it’s correct — I just don’t think the market has finished deciding whether it’s willing to pay the setup tax to get there.
