# Build the developer preview.

> Build the Overtalk developer preview on Linux, choose local models for speech and notes, and check your first recording.

Source: https://overtalk.io/setup.html · Facts checked: 2026-10-20

For developers comfortable with Rust and native dependencies. Linux is the tested desktop path. Windows and macOS builds and capture support still need validation before a general release.

**Just exploring?** [Try the sample meeting](https://overtalk.io/index.html#sample) with no installation.

**Not a developer?** There is no installer yet. Packages for Linux, Windows and macOS are being prepared; [see how to hear when they ship](https://overtalk.io/pricing.html#get-started).

## What you need

- **Linux with a PulseAudio-compatible sound server**, which includes PipeWire with pipewire-pulse. Builds are tested on Ubuntu 24.04. Overtalk records the call by reading your output device's monitor source and never changes your default devices, profiles or volume.

- **An x86-64 processor.** A source build targets your own processor; the coming packages need x86-64-v3 (Intel Haswell or AMD Zen, 2013 onwards). An NVIDIA GPU is optional: CUDA and Vulkan are build options.

- **About 1.7 to 3.3 GB for models.** Recording, transcription and notes need two downloads, each pinned and checksummed: Parakeet TDT 0.6B v3 for speech (670 MB, 25 European languages) and a notes model of 1.0 to 2.6 GB (Gemma 3 1B, Qwen 3.5 2B or Qwen 3.5 4B). Starting the app downloads nothing.

- **bubblewrap**, which isolates the model workers, and FFmpeg.

A .deb for Ubuntu 24.04 or newer and Debian 13 or newer, and an AppImage, are being prepared for the first release. Until then, build from source.

## Prepare the desktop app

Get the source from the [Overtalk repository](https://github.com/judeam/overtalk).

Install Rust, Node.js, pnpm and your platform's native audio/UI libraries. Follow the repository's [build prerequisites](https://github.com/judeam/overtalk/blob/main/docs/BUILDING.md) to configure FFmpeg, inference libraries and sidecars before compiling.

The native build uses a Slint interface and a Rust service. From a prepared checkout:

`cargo build --release -p overtalk-native cargo build --release -p sonefold --bin sonefold-service cargo run --release -p overtalk-native`

The full [native build guide](https://github.com/judeam/overtalk/blob/main/docs/NATIVE-UI.md) covers the additional worker sidecars, platform requirements and installation steps. These commands are not a replacement for those prerequisites.

## Choose where processing happens

For local speech and notes, download and configure models suited to your machine. The built-in speech engine needs no Python. The optional local speech worker adds Qwen3-ASR 1.7B, with 30 languages including Arabic, Chinese, Hindi and Japanese, and speaker separation; it runs in its own Python environment. See the [speech setup guide](https://github.com/judeam/overtalk/blob/main/speech-worker/README.md).

You can instead configure a supported external notes provider. That sends the selected content to that provider, subject to its terms. Cloud access is optional.

## Check your first recording

1. Open the app and use the sound check to confirm your microphone and call-audio source.

1. Make a short recording with everyone's knowledge, or import an audio file.

1. Play it back, review the transcript and correct speaker labels if needed.

1. Generate notes with your selected model and check important decisions against their source.

Your first useful record is the goal. Projects, briefs, remembered voices and connected tools are all in the same app.

## Connected tools

Connect Linear, Jira and Confluence, Google or Microsoft 365 in Settings → Connections; follow-ups are sent only after you review the draft. Linear and Atlassian use account sign-in. Google and Microsoft need a registered desktop client in this preview; see [account linking](https://github.com/judeam/overtalk/blob/main/docs/CONNECTION-ACCOUNT-LINKING.md). You can always copy or export the notes instead: PDF, DOCX, Markdown, text and JSON.

[Projects, filing and briefs](https://overtalk.io/pricing.html#who)

## AI assistants and the command line

`overtalk-cli` reads your meetings from a terminal and is an MCP server for Claude Code, Claude Desktop, Codex, Cursor and VS Code. It is off until you turn it on.

1. In Overtalk, open Settings → Automation → Connections → Configure, turn on **Command line access** or **MCP access**, choose what it may read, and save. Recording control and note editing are separate switches.

1. Build it: `cargo build --release -p overtalk-cli`. The Linux packages install it as `overtalk-cli`.

1. Connect your assistant: `overtalk-cli setup claude-code` prints the configuration, and `--write` adds it for you, keeping a backup. The same works for `claude-desktop`, `codex`, `cursor` and `vscode`. `overtalk-cli skill install` adds the Overtalk skill to Claude Code.

Overtalk proves who it is before the tool sends its access token, so a program that takes the port while Overtalk is closed learns nothing. An assistant may send what it reads to its own model provider; grant only what you want it to see. [Every command and grant](https://github.com/judeam/overtalk/blob/main/integrations/README.md)

## Keep control of the record

Store recordings somewhere you control, protect the device and keep appropriate backups. Review the app's privacy and retention settings before handling sensitive conversations.

[Read how data moves through Overtalk](https://overtalk.io/privacy.html)