Parrot: Open Source AI Meeting Recorder for Mac

A close read of Parrot's design: bot-free capture, a locally indexed knowledge base, per-call profiles, swappable models with visible costs, and the limits engineers should know before adopting it.

Binary Buddies · 5 min read · 5 views

Most AI meeting tools solve the wrong half of the problem. They join your call as a visible bot, record it, upload the audio, and hand you a summary an hour later. By then the useful moment has passed: the prospect has already asked about volume pricing, or the candidate has already given a vague answer you should have probed.

Parrot targets the moment itself. It is a Mac meeting recorder with a live assistant that notices when a call starts and suggests answers from your own documents while you are still talking. It also keeps every call so you can query it later. The project's own tagline is 'Help during the call. Not after it.'

It is free and open source under GPL-3.0. It ships as a 25 MB download for macOS 14 or later on Apple Silicon, or you can build it from source. This article covers what the project page says about how it works, where the data goes, and where it does not fit.

Recording without a bot in the participant list

Parrot does not join the meeting. According to the project, it records what your Mac already plays and hears, so nothing appears in the participant list. That one design choice removes the most common objection to AI note-takers: an unannounced third attendee called 'Notetaker' that makes the other side wary before you have said anything.

Because capture happens at the machine level rather than through a meeting-platform integration, it works with anything that makes sound. The page lists Google Meet, Zoom, Teams and a phone on speaker. If your Mac can hear it, Parrot can. When a call app starts, Parrot asks whether to record, so you get a prompt rather than silent capture.

The flip side is that nobody else on the call sees a recording indicator. That has consequences, covered in the final section. The project page does not describe the macOS audio APIs it uses, so check the source for those details.

What stays on the Mac and what leaves it

The privacy model is the most concrete part of the project page, so it is worth reading closely. Three things run locally by default:

  • transcription
  • speaker detection
  • indexing of your documents

Cloud transcription happens only if you switch it on. The app needs no account, sends no telemetry, and the build is signed and notarized.

The interesting nuance is the assistant. If you pick a hosted model such as Claude, the app states plainly what crosses the wire: 'Transcript text, to Anthropic. Never the audio.' Each model choice is labelled with exactly what it sends. That is the right level of disclosure for engineers who have to explain a tool to a security reviewer. The answer is not 'nothing leaves', but it is precise.

For calls where even transcript text must stay put, you can set a profile to on-device only, and nothing about those calls leaves your Mac. A Pause button on the call screen stops all sending and all spending while it is active. That covers the moment someone says 'let me share something confidential'.

A knowledge base that is indexed locally

The live suggestions are only as good as what the assistant can ground them in. Parrot's answer is a local knowledge base. The project's phrase is to brief it like a new teammate. You drop in a pricing sheet, an FAQ or a playbook as PDF, plain text or markdown. Parrot chunks and embeds those files on the Mac using Apple's own language models. The page says nothing is uploaded, ever.

Retrieval is cross-lingual. An English question can find the answer in a Turkish or Spanish document, which matters for teams whose collateral exists in more than one language.

Every suggestion card names its source: either your document or the words 'general knowledge'. You decide whether the assistant may answer beyond your files at all. For a sales call that is the difference between quoting your actual pricing and confidently improvising one.

Organisation is folder-based:

  • Group documents into folders, such as 'Acme basics' or 'the Northwind deal'.
  • Give each folder one 'Use for' setting: all call types, a selection, or Off to pause it.
  • Override the setting for an individual document if needed.
  • Add an 'about' line to a folder, and use Search documents to find files by name.

On top of the documents sit two kinds of instruction. Coaching instructions are standing guidance applied to every call, such as 'keep answers short' or 'always offer three price options'. A pre-call brief is a line or two typed before you hit record, so the assistant knows who you are talking to from the first second. When you meet the same people again, it starts with what is still open from last time.

Profiles: changing what the assistant watches for

A sales discovery call and a job interview need different things from a live assistant. Parrot handles this with profiles. A profile decides what the assistant watches for during the call and how the report is written afterwards. There are eight built-in profiles, an investor pitch among them, and you can build your own.

During the call, each profile has its own card types and gauges. The sales profile surfaces objections ('price is above what they budgeted'), buying signals ('asked how fast onboarding is') and prompts for questions you forgot to ask ('who signs off on budget?'). An interview profile instead looks for red flags and strong answers.

After the call, the profile shapes the report: which sections it has, in your wording, which one holds the promises made, and whether coaching is on. Sales reports ask about budget and decision-makers; interviews get a scorecard.

Profiles are portable. You can export one for a colleague or import theirs, and Parrot shows every change before applying it. The exported file never carries meetings, documents, names or keys, so sharing a profile does not leak a customer list.

Connecting a calendar lets Parrot choose the profile automatically. An event titled 'Interview: Jane' picks the Interview profile.

Choosing the model and seeing the bill

Parrot does not lock you to a provider. You can run the AI locally with Ollama, or bring an API key for Claude, OpenAI, Gemini and others. The project describes Claude as the best quality for live cards, and it needs a key. A separate TypeSafe AI key handles fast document answers, returning matching excerpts from your files within a second of a question being asked.

Cost is visible per meeting. Each one shows the model, tokens, number of calls and minutes. Local features show $0.00. That makes the cloud-versus-local decision concrete: you can see what a week of sales calls actually costs on a hosted model.

Two controls shape spend and latency:

  • Pace. 'Relaxed' fits free API tiers. 'Fast' is for calls where every second counts.
  • Context window. You also choose how much of the conversation each request carries, which drives token cost directly.

'Ask Parrot', the feature for querying past calls, uses your reports model by default. In Settings under Assistant you can point it at Ollama instead, which keeps it free and on the Mac.

Transcription and language handling

Parrot auto-detects the language of the call. You can also pin a language in Settings, or set one for a specific call under Start recording.

Two models do the work. For European languages, Parakeet v3 is a faster free option covering 25 of them, but not Turkish. Everything else goes to Whisper. If a call turns out to be in a different language than expected, a banner says so and offers Switch or Keep, so a mid-call language change does not silently produce garbage.

The assistant's cards and the post-call report are written in the language of the call. The project notes that Claude handles the most languages well, which is worth weighing if your calls are not in English. The page lists English, Turkish, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Arabic, Hindi, Chinese, Japanese and Korean 'and dozens more'.

Open source as a procurement argument

For a tool that listens to every call, the GPL-3.0 licence is more than a philosophical point. A team can read exactly what the app captures and sends, rather than trusting a privacy page. It can build from source and pin a version, and it can keep using the tool if the maintainers change direction. Combined with no account and no telemetry, that is a short conversation with a security team compared with most hosted note-takers.

GPL-3.0 is a copyleft licence. Using Parrot internally is straightforward, but if you plan to modify it and distribute the result, for example as part of a commercial product, your derivative carries GPL obligations. Have whoever handles licensing read it before you fork.

Where it does not fit, and what to watch for

The hardware floor is firm. Parrot requires macOS 14 or later on Apple Silicon. Intel Macs, Windows and Linux machines are out, so a mixed-OS team cannot standardise on it.

No bot in the participant list also means no automatic notice to the other side. Recording-consent rules vary by jurisdiction. Some require every party to agree, and an invisible recorder makes it easy to forget to ask. Treat disclosure as your process, not the tool's.

'Audio stays on your Mac' is true by default but conditional in practice. Pick Claude or another hosted model and transcript text goes to that provider. Enable cloud transcription and that changes too. For regulated calls, use an on-device-only profile and make sure the right calendar events map to it, rather than relying on people to remember the Pause button.

Finally, local-only is a trade. The project itself says Claude writes the sharpest cards and handles the most languages well. Running everything through Ollama costs $0.00 and sends nothing, but card quality and speed depend on what your machine can run. For Turkish calls, Parakeet's speed advantage does not apply. Pilot it on real calls in your own languages before rolling it out.

  • #AI meeting recorder
  • #GPL-3.0
  • #Mac meeting assistant
  • #Ollama
  • #Open Source
  • #local transcription

Comments

Add a comment
Sign in with Google to join the conversation