touri-ai/touri

touri is a local-first desktop app for macOS, Linux and Windows for running persistent AI teammates ("bots") on your own
Codex, Claude Code or OpenCode CLI. Everything runs on your computer: your conversations, memory, routines
and secrets never leave the machine unless you explicitly configure a remote.
Built with Tauri v2 (Rust shell), a Bun sidecar and a React UI.
## Features
Persistent bots with their own instructions, memory, skills and full branching chat history.
Codex and Claude Code runtimes, with approval modes (on-request, workspace-write, full).
Routines: cron-scheduled bot runs with overlap and catch-up control.
Projects: connect a local repo, run supervised processes (databases, servers) and preview them in-app.
Agent browser with human take-over, and opt-in computer use with a visible indicator and kill switch.
MCP gateway with per-thread tokens and approval prompts.
Bot machines: Apple container on macOS; rootless Podman or Docker containers on Linux.
Importer for data exported by an earlier setup (see below).
API models with your own key, model discovery and a connection test in Settings.
### Your own API key
Open Settings → API models (Configurações → Modelos por API), choose a provider, paste its API key,
and load the model list or enter a model ID. Test connection sends a short real message and may use a
small amount of provider credit. Save the model, then select it when creating a bot or in the bot settings.
Presets cover OpenRouter, DeepSeek, Z.ai/GLM, Moonshot/Kimi and Groq. A custom compatible API needs its
base URL, including its API version path. Use HTTPS; HTTP is accepted only for loopback servers.
Choose a chat model that supports tools. API models run through OpenCode with the same folder permissions,
tools and approval bridge, locally or in a bot's isolated machine; the cloud runtime does not support them.
If the executor is missing, the setup screen links to installation instructions and offers a recheck.
The selected provider bills usage directly. Keys use the existing system secret store, or the encrypted
secret store in headless Linux server mode. Only model metadata enters SQLite. Removing a saved model
deletes its key; bots using it ask you to choose another model instead of falling back to another account.
## Privacy
Local-first: data lives in SQLite in the OS application data directory.
No telemetry and no account required.
Desktop credentials use macOS Keychain, Linux Secret Service or Windows Credential Manager, never the database, logs or exports.
Optional cloud hosting is a separate, opt-in protocol (docs/protocol/cloud-v1.md).
## Platforms
macOS (Apple Silicon): pre-release, macOS 14 or newer.
Linux (x8664/arm64): supported. Packages and desktop onboarding have been tested on x8664;
the arm64 sidecar is cross-built, with native arm64 packaging/runtime validation still pending.
Windows (x64): unsigned MSI and EXE installers; see the platform runbook for remaining limitations.
## Install (GitHub Releases)
Release v0.3.0 provides packages for all three platforms. See the release notes in Portuguese and English.
The links below target the v0.3.0 assets and become available when that release is published.