Patum for engineering and product teams
Meetings that answer from your notes and your code.
Patum records stand-ups, design reviews and incident calls on your Mac, transcribes them there and writes notes that keep each decision with its reason. When someone asks what the code does today, Live context answers during the call from your repositories and cites the symbol and the line. With Pro, Claude Code, Codex, Cursor and Claude Desktop read your meetings through a local MCP server.
// Retry schedules a failed upload again: up to three
// more tries, the wait doubling from 2s to at most 30s.
// A 4xx other than 429 is not retried.
func (q *Queue) Retry(ctx context.Context, job Job, err error) error {
if job.Attempts >= maxRetries || !retryable(err) {
return q.fail(ctx, job, err)
}
wait := min(baseDelay<<job.Attempts, maxDelay)
return q.schedule(ctx, job, wait)
}
Where engineering notes go wrong
A design review settles something at minute 18 and the note says "agreed to change retries". It doesn't say why, who changes it, or by when, and a month later the same argument starts again. Halfway through a call someone asks what the code does today, and the answer is a guess or "I'll check after". Your coding agent knows none of it: it can grep a notes folder, reading whole files to find two sentences.
A cloud notetaker fixes the typing by sending the call, with your system names and customers' names in it, to a vendor that keeps the transcript. For a lot of internal design talk, that is a harder sell than the notes are worth.
Patum does the listening, the writing and the answering on your Mac.
During the call: answers from your code
Add your repositories in Settings › Live context › Code repositories. Patum indexes them on your Mac: functions, methods, types with their fields and signatures, and Markdown sections, each with its doc comment. It skips what git ignores, hidden files such as .env, dependencies, build output, lock files and files over 512 KB. It reads a folder through a read-only bookmark and never writes to it.
In the recording window, pick Code to ask about the code alone, or turn on Also Search Code next to a notes project. When someone asks "what does the upload queue do on a failure?", Patum looks for names first (an identifier in the question, or the same said as words), then by meaning and keywords. The answer names the function or file, explains it without pasting code, and cites it as [c41]: a click opens the file at its line in VS Code, Cursor or Zed. When the code doesn't hold the answer, Patum says so: "Not in your code or the call."
Turn on Answer Questions Heard in the Call and questions like Jon's are answered as they are asked, without anyone typing. The index, the search and the answer all run on your Mac. Code excerpts reach a cloud provider only if you chose one (Pro) and also turned on Include code excerpts; without both, a question about code is answered on the Mac or not at all.
Live context over your notes and your code is part of Free.
After the call: decisions with their reasons
The note keeps what an engineering meeting is for: decisions with their reasons, owners and dates, open questions, risks, and the exact names of systems and tickets. When nobody named an owner, the note says so ("Load test: no owner was named"), because Patum takes out owners and dates the model inferred. Every fact links to the second it was said.
Templates follow the kind of meeting, and the default, Detect from the meeting, picks one for you:
- Stand-up: per person, what's done, what's next and what's blocked, then the team's risks.
- Project sync: milestones and their status, risks, decisions, actions.
- Retrospective: what went well, what didn't, and the agreed changes with their owners.
- 1:1: commitments both ways and threads to pick up next time, linked to the previous 1:1.
- General meeting, for everything else, design reviews and incident calls included.
The note filed to your vault also has a Details section: every fact Patum read from the transcript, by topic, each with its time and speaker, and every action in one table.
Your coding agent reads your meetings, locally
With Pro, turn on Local MCP server in Settings › Integrations. Your AI tool starts a small helper inside Patum over stdio, and the helper passes each call to the app over a local socket that only it may use. Nothing listens on the network.
Add it to Claude Code with:
claude mcp add --scope user patum -- '/Applications/Patum.app/Contents/Helpers/patum-mcp'
Copy Setup in Settings gives the same for Claude Desktop, Cursor and Codex. The tools only read:
| Tool | What it returns |
|---|---|
search |
up to 8 cited excerpts from notes and meetings, ranked together |
ask |
an answer from your library, citing notes and meetings |
list_meetings |
recent meetings, by date or project |
get_meeting |
one meeting's note, and the transcript lines about a query |
list_projects |
your projects and how many meetings each has |
An agent gets a few excerpts per call instead of whole files. Each starts with a line that says where it came from:
[m412] “Design review: upload retries” · 2026-11-03 · meeting 8F2C1A · Summary: Decisions · score 0.031
It reads only meetings in your projects and clients, as Settings › Integrations › What agents can read allows. People's pages, the Journal, meetings without a project, Recently Deleted, audio, voice samples and code folders are never read.
Patum's side stays on your Mac. Your AI tool is another matter: Claude Code, Codex, Cursor and Claude Desktop send what they read to their own model, under their own terms. A cloud connector works differently: with Granola's MCP, for example, the meeting data itself lives on Granola's servers (Granola MCP).
Markdown you own, filed by project
Each note is a Markdown file under its project, Projects/Platform/2026-11-03 Design review — upload retries.md, with front matter you can query and a patum://rec/… link that opens the meeting at its moment. Without Pro it goes into Patum's own vault; with Pro, into your Obsidian vault, inside a folder named Patum. Edit anything: your edits are never overwritten, and a later version goes beside your note as "(Patum update)". The Obsidian guide shows the whole layout.
Pro also adds:
- Ask across your library, one project or one person, at any time, with each answer's sources.
- A regular summary of a project every week or month, each fact still pointing to its moment.
- Vocabulary, so system names and jargon come out right in transcripts, speaker names and notes.
- Screen moments: a picture of the call's window when what it shows changes, such as the next slide of a design review, with the text read from it.
For a team that can't use cloud notetakers
In Patum for teams, each Mac records, transcribes and summarizes its own calls, and audio is never shared. A note reaches colleagues only when its author publishes it from the Outbox, encrypted with keys made on the team's Macs; the relay that passes it along can't read it on its own. Admins can keep work meetings off cloud AI providers and require the recording notice and Touch ID. When someone leaves, their work meetings are deleted from their Macs and their personal ones stay.
No outside audit has been done. What isn't proven yet is on the Team page.
What it costs, for one engineer or a team
Free records, transcribes, writes the checked notes and answers from your notes and code during the call, with no minute limits. A plan adds the rest.
One engineer
Pro
$79 once, after 14 days free
The MCP server for Claude Code and Cursor, Ask across your library or a project, your own Obsidian vault, and Vocabulary for your system names. On up to three of your Macs.
A team of 3 to 50 people
Team
$16 a seat a month, billed yearly
Pro for everyone, published notes encrypted between your Macs, and Search, Ask and Live context over the team's notes.
A company whose IT runs its Macs
Enterprise
No list price
Your company's AI server for work meetings, your own relay inside your network, Macs prepared through MDM, and keys for Macs that never go online.
A team of ten, for a year
Ten seats of Team cost $1,920 a year, billed yearly (Team's prices). The consultants' page sets that year beside Granola's, Fathom's and Otter's, each from its own pricing page.
Granola and Fathom cost a little less per seat, and Granola and Otter also run on Windows and phones, which Patum doesn't. With Patum Team, Pro and its MCP server come with every seat, on up to 3 Macs each, and the audio of a design review never leaves the Mac that recorded it. A team can try it free for 30 days first.
What it costs your Mac
Live transcription runs on the Neural Engine; answers and the note use the GPU. Patum loads its models only for the recording and the note, and unloads them when idle. Expect it to compete with a heavy build during a call. Patum needs 16 GB of memory or more, and Macs with 8 GB aren't supported. With Pro, a cloud provider with your own key can write the notes and answers in place of the Mac.
The models, each downloaded once from Hugging Face:
- speech: Parakeet (English and 24 other European languages) or Whisper (99 languages);
- speakers: NVIDIA's Nemotron diarization, and a voice model that recognizes people you've named;
- notes and answers: Qwen3 models on MLX, the 9B for notes on Macs with more than 16 GB and the 4B otherwise.
How Patum works lists every model with its license and size.
Setting it up
- Download Patum and open it. Setup asks for the microphone and call audio, and downloads the models once, showing their size first.
- Add your repositories in Settings › Live context › Code repositories. The first index of a large repository takes minutes; later ones read only what changed.
- Make a project for the team, such as Platform. In its Overview, Choose a calendar series… files that series' meetings there from then on.
- In your next call, pick the project in the recording window and turn on Also Search Code.
- With Pro, turn on Local MCP server in Settings › Integrations and use Copy Setup for your AI tool.
What Patum doesn't do
- It runs only on Macs with Apple silicon, macOS 26 or later and 16 GB of memory or more. There is no Linux, Windows or iPhone version.
- It isn't open source, so you can't read or audit its code yourself. If you need that, choose an open-source app.
- The local 4B and 9B models write plainer prose than frontier models. With Pro you can use your own OpenAI, Anthropic, Google Gemini, xAI or OpenRouter key for notes; the transcript then goes to that provider.
- It reads code without a compiler, so unusually formatted code can split badly, and an answer reads at most six excerpts: there is no call graph beyond them.
- It doesn't push notes to Jira, Linear, Notion or Slack. MCP and Shortcuts are the way out.
- It doesn't split one recording into several notes.
Questions engineers ask
Local models are weak. Are the notes any good?
Patum doesn't lean on the model's prose. It reads the transcript into facts, writes the note from them, then checks owners, dates, numbers and who said what against the transcript and drops what it can't back. The prose is plainer than a frontier model's; every fact is traceable.
Does it know our system names?
During the call, Live context finds names from your code index, written as code or said as words ("upload queue" finds UploadQueue). For transcripts, Vocabulary (Pro) takes your system names, products and people and uses them in transcription, speaker names and notes.
What about secrets in our code?
Patum skips hidden files such as .env, what git ignores, dependencies, build output and lock files. A secret written inside a source file is indexed like any other line, and the index stays on your Mac. Code goes to a cloud provider only if you chose one (Pro) and turned on Include code excerpts.
Can my team see my notes?
Only if you're in a team and publish a note. Without Team, nothing leaves your Mac.
Does it work in Slack huddles and Google Meet?
Yes. Patum notices calls in Slack, Zoom, Teams, Webex, Discord, FaceTime and others, and in a browser tab that uses the microphone, such as Meet in Chrome, Safari, Arc, Edge, Firefox, Brave, Vivaldi, Opera, Chromium, Dia, Comet, Orion or Zen. A browser is never recorded without asking; Add an app… in Settings › Recording › Call apps adds any other.
Can we use our own model server?
In Enterprise, an admin can name your company's AI server for work meetings: LM Studio, vLLM, Ollama or another OpenAI-compatible server, so Macs with less memory don't have to write the team's notes themselves. On your own, Pro can use a model in Ollama on your Mac, or a cloud model with your own OpenAI, Anthropic, Google Gemini, xAI or OpenRouter key.
Can my AI tool change anything in Patum?
No. The MCP tools only read, and never see code folders.
Try it on your next design review.
Free, with no account and no minute limits. Pro is free for 14 days, no card, then $79 once. For your team, the Team plan is $16 a seat a month, billed yearly.