Takku

Set up Takku

Install the app, allow the recording, then choose a folder for the notes and who writes the summary. You do it once.

You need a Mac with Apple Silicon and macOS 15 or later. Nothing else to install: the notes are Markdown files, and any app opens them.

Install

  1. Download Takku, open it and drag Takku into Applications. With Homebrew: brew install --cask mameli/takku/takku
  2. Open Takku. Apple has notarized it, so it opens like any other app.

Takku lives in the menu bar, not in the Dock.

First setup

  1. The first time you open Takku, a welcome window walks you through the steps below. You can do each one later in Settings too.
  2. Choose where the notes go: Use Documents/Takku, or any folder, your Obsidian vault too. Takku writes each meeting in its Meetings folder.
  3. Allow the microphone and System Audio Recording Only when macOS asks.
  4. For screenshots during meetings (⌃⌥⌘S), allow Screen Recording too. It’s optional and takes effect the next time Takku opens.
  5. Download the transcription model, about 630 MB. macOS prepares it for your Mac once, in a few minutes. Skip it and Takku does it during your first meeting.
  6. Choose who writes the summary: buy or paste a Takku Credit key right in the welcome window, click Use Apple Intelligence (on macOS 26 or later), or open Settings, Summary to use one of the ready-made profiles (Ollama, Mistral, TensorX, OpenRouter) or add your own. API keys are kept in the macOS Keychain.
  7. If you want Takku to start with your Mac, turn on Open at login.

Profiles to start from

Apple Intelligence, on your Mac
Built into macOS 26 or later: free, nothing to set up and nothing leaves your Mac. Its model is small, so long meetings take a few minutes and summaries are shorter than with a large model.
Ollama, llama.cpp or LM Studio, on your Mac
In the ready-made Ollama profile, pull a model (for example ollama pull hf.co/ggml-org/gemma-4-E4B-it-GGUF:Q4_0) and start Ollama with OLLAMA_CONTEXT_LENGTH=32768. With llama.cpp, run llama-server -hf ggml-org/gemma-4-E4B-it-GGUF:Q4_0 -c 32768 and use http://localhost:8080/v1.
Mistral (France) or Regolo.ai (Italy)
Paste your API key into the ready-made Mistral profile (mistral-medium-latest), or add Regolo.ai with base URL https://api.regolo.ai/v1 and model gemma4-31b.
OpenRouter (United States)
Paste your API key into the ready-made OpenRouter profile and pick any model from its catalogue. OpenRouter is outside the EU and passes requests to each model’s provider: know where your meetings go.
A coding agent you already pay for
With Codex, GitHub Copilot, Claude Code, OpenCode or Pi installed and signed in on your Mac, the + in Settings, Summary adds it as a profile: the summary uses your plan, with no key to paste. Takku runs it with no tools and none of your setup, so a sentence said in a meeting can’t make it run a command or read a file. Claude Code at your own risk: Anthropic’s terms are unclear on other apps running it with your plan.

When adding a profile of your own, set Max context to the model’s real limit (for a local server, the -c you started it with). To keep only the transcript, choose Transcript in the Takku menu.

Takku doesn’t check where a provider processes your data: read its terms, and your company’s policy, before using it for real meetings.

Reasoning and extra parameters

Reasoning sets how much the model thinks before answering. Model default sends nothing and is the safe choice. OpenRouter adapts every level to the model; other providers get OpenAI’s reasoning_effort, which not every model accepts.

Extra parameters is an optional JSON object merged into every request, overriding Takku’s keys when a model needs custom parameters (see your provider’s docs). Two examples on TensorX:

DeepSeek V4
{"chat_template_kwargs": {"thinking": true}} turns reasoning on.
Qwen 3.8
{"reasoning_effort": "xhigh"} enables maximum reasoning (it refuses Off and High).

Press Test connection after changing either setting: it sends a test request with the same reasoning and extra parameters as a summary.

Updating

From 1.17 Takku updates itself: when a new version is out, its menu says so; choose it, then Install Update. Takku never restarts in the middle of a meeting. From an older version, run brew upgrade --cask takku, or quit Takku and replace it in Applications: step by step. Takku keeps your settings, profiles, API keys and models.

Coming from a 0.x version: from 1.0 Takku carries an Apple Developer ID signature, so macOS asks once more for the microphone, the system audio and the calendars, and for your password to let Takku read your saved API keys (choose Always Allow).