Can I use my own OpenAI API key instead of a subscription?

You can. In Narada, Settings → AI has a mode called OpenAI API key; paste one in, grant cloud-AI consent, and every summary, follow-up extraction and chat answer runs against your key and your billing account. Narada takes no cut and needs no subscription for it.

It is one of five routes, and they are not marketing tiers. They are different network paths with different bills and different privacy consequences.

The five routes

Recording, on-device transcription, the files in your folder and labels work identically in all five. The mode decides who writes the summaries — and, indirectly, what the Claude Desktop connector has to read: it always lists meetings and returns transcripts, but in Recorder only there are no summaries, no follow-up ledger, and no transcript search index, because indexing runs on the same route as the AI.

Who each suits: Recorder only if you want an archive and will read it yourself. On-device AI if nothing may leave the machine and you have an Apple Silicon Mac with the disk and memory for it. Your own key if you have an OpenAI account and prefer a metered bill. Custom endpoint if you run LM Studio or Ollama and want a better local model. Pro AI if you want none of that to be your problem.

What a meeting actually costs on your own key

OpenAI’s pricing page lists gpt-5-mini — the model Narada uses for analysis and chat synthesis — at $0.25 per million input tokens and $2.00 per million output, and text-embedding-3-small at $0.02 per million.

The other half I measured rather than guessed. Across a real archive — 167 recordings of 15 minutes or more, tokenised with the encoding the GPT-5 family uses — transcripts run at a median of 341 tokens per minute of meeting, so an hour-long meeting is about 20,000 tokens of text. Script matters: mostly-Cyrillic transcripts came in at 376 tokens a minute, mostly-Latin ones at 230. Anyone quoting a flat “4 characters per token” is quoting an English rule.

The transcript is not read once. Analysis is chunked at up to 16,000 characters a chunk, then synthesised, and follow-up extraction reads it again. So:

This is an illustration, and the multiplier is the soft part of it. The prices are published and the transcript length is measured; “about three times” is an estimate of the passes the pipeline makes, and a meeting full of commitments makes more.

Chat is cheaper than it feels. The app’s own requirements text puts its prompts at “up to ~25k tokens” — under a cent of input, so two hundred questions a month is near $1.50. Twenty hour-long meetings plus daily chat comes to well under $6.99. That is the honest headline, and it does not favour the subscription.

When your own key costs more than $6.99

Three ways, and the first is the real one.

Bulk re-analysis. The meeting screen has Regenerate and Full Regenerate; Follow-ups has Rebuild. Point those at a 300-meeting archive and you have made several hundred calls in an afternoon — around $6 on the arithmetic above, with no dialog quoting a price and no spend meter anywhere in the app. Set a monthly limit in your OpenAI dashboard; Narada will not do it for you.

Long meetings. The cost is per minute of speech. A three-hour workshop is three times an hour-long call, and a week of those adds up in a way a flat subscription does not.

No shared prompt cache. Pro AI goes through Narada’s own proxy, which exists partly to cache prompts centrally — the project’s decision log records the hit ratio going from 12% to 78%. A direct key talks to api.openai.com and shares none of that, so identical work can cost more tokens.

One line is missing from this bill, and it is normally the biggest: turning an hour of audio into text costs zero tokens in every mode, because transcription runs on your Mac.

Where the key is stored

In the app’s own preferences file inside its sandbox container — ~/Library/Containers/stepanok.com.Meetings-Recorder/Data/Library/Preferences/, under openai_api_key. Not in the macOS Keychain. Other apps cannot read that container, but it is a plist on your disk and anything running as you can.

A Test button validates the key against OpenAI before you rely on it; Remove clears it. The key belongs to the machine rather than to the archive, so move your recordings folder to another Mac and the meetings come with it — the key does not.

The custom-endpoint route, and the one knob you must set

The fifth mode points Narada at your own OpenAI-compatible server. Settings spells out its requirements, and they are hard ones:

• Three endpoints: /v1/chat/completions, /v1/embeddings, /v1/models. LM Studio, Ollama, llama.cpp’s server and vLLM all qualify.
• A chat model of 7B parameters or more. The app names what it was tested against — gpt-oss-20b, Qwen2.5-14B, Llama-3.1-8B — and says smaller ones “miss multi-meeting context and citations”.
• 32k context, set by hand. This is the one that catches people, and the app says so outright: "This is the only knob you have to set by hand." Prompts reach about 25k tokens, and a default 4k context returns empty replies. LM Studio: model card → Configure → Context Length → 32768 → Reload. Ollama: OLLAMA_CONTEXT_LENGTH=32768, or PARAMETER num_ctx 32768 in the Modelfile.
• An embedding model loaded alongside it, if you want chat and search over the archive. Summaries work without one; chat does not. Any dimension is fine — the app learns it on first use.

Switching modes empties chat until it re-indexes

Stored vectors are keyed to the model that made them. Cloud mode embeds with text-embedding-3-small at 1536 dimensions; on-device mode uses a small multilingual model at 384. Those are not comparable numbers, so a switch invalidates the index and Narada starts a full rebuild in the background.

Recordings, transcripts and summaries are untouched, and the rebuild is cheap — re-embedding a 300-meeting archive is roughly a dime. But chat and search return nothing useful until it finishes, so switch once, deliberately, rather than flipping between modes to compare them.

What your own key does not get you

• No cost display. The app never shows what a meeting cost or what you have spent. That lives in OpenAI’s dashboard, and so does the only spending cap.
• No other providers. One key field, OpenAI-shaped. No Anthropic key, no Gemini key; a gateway behind the custom-endpoint mode is the workaround, and it has to serve embeddings too, which many do not.
• No cheaper transcription, because it was never billed. You cannot bring a key for speech-to-text. Transcription is the on-device model in every mode, across 25 European languages, with published word error rates from about 4% on Italian to about 37% on Greek.
• No named speakers. The call’s audio is split into Speaker 1, Speaker 2 and so on, and the app refuses to guess who they are. Your lines are attributed to you only because your microphone is a separate file, so a mic-only or in-person recording gets no separation at all.
• No guarantee the recording is whole. macOS audio can fail mid-capture. Narada compares each finished recording against itself and marks it Incomplete with a reason; it cannot recover audio that was never written.

And recording other people is your legal responsibility in all five modes. The app shows a reminder on your own screen and tells nobody in the call. Rules differ by country and by state. Say it out loud at the start.

Sources

• OpenAI API pricing — gpt-5-mini at $0.25/$2.00 per million tokens, text-embedding-3-small at $0.02 per million: https://developers.openai.com/api/docs/pricing
• Narada’s own App Store listing — “Pro Monthly $6.99”, “Pro Annual $69.99”, “Requires macOS 15.6 or later”: https://apps.apple.com/us/app/narada-meetings-recorder/id6766092471
• Transcript token density measured on 4 August 2026 across 167 recordings of 15 minutes or longer in a live Narada archive, tokenised with the o200k_base encoding used by the GPT-4o and GPT-5 families: median 341 tokens per minute, 376 mostly-Cyrillic, 230 mostly-Latin.
• Model routing, the 16,000-character analysis chunk, the local-server requirements and the preferences key: Narada 1.2's own source and Settings text.
• Word error rates per language: the FLEURS figures published in the documentation of FluidAudio 0.13.4, the on-device transcription package Narada ships. https://raw.githubusercontent.com/FluidInference/FluidAudio/main/Documentation/Benchmarks.md

Checked 4 August 2026.