thunderbird/thunderbolt

What it solves

Thunderbolt addresses the problem of vendor lock-in and data control in AI chat applications. It is an open-source, cross-platform AI client that lets you choose which models you use (frontier, local, or on-prem) and where your data lives. You can deploy it entirely on your own infrastructure, which matters for enterprises that need privacy and compliance.

How it works

Thunderbolt is a client application available on web, iOS, Android, Mac, Linux, and Windows. It connects to a backend you can self-host using Docker Compose, Kubernetes, or Pulumi on AWS. The client supports OpenAI-compatible model providers, as well as local inference through Ollama or llama.cpp. It also ships system-managed models (e.g., Opus 5.5 via Anthropic, and confidential models in Tinfoil enclaves) if you provide the corresponding API keys.

The app is designed as an extensible workspace: you can define projects (which share instructions across chats), skills (reusable instruction bundles), widgets (interactive components the model embeds), HTML artifacts (sandboxed pages), and voice mode. It includes a command palette (Cmd/Ctrl+K) for searching chats and messages, and supports MCP (Model Context Protocol) servers to connect external tools. A separate CLI (thunderbolt) acts as a terminal coding agent and bridges to the app via ACP/MCP.

Data is stored locally-first in SQLite, with optional multi-device sync (via PowerSync) and end-to-end encryption. Attachments are never stored server-side. The backend handles authentication, rate limiting, and a universal proxy for model requests.

Who it's for

Thunderbolt is primarily aimed at enterprises that want to deploy an AI assistant on-premises, with full control over models and data. It is also suitable for individual developers who want to self-host and use local models. The project is early-stage and under active development, so it may not be production-ready for all use cases yet.

Highlights

• Cross-platform: web, iOS, Android, Mac, Linux, Windows.
• Model-agnostic: works with frontier APIs, local (Ollama, llama.cpp), or on-prem models.
• Self-hostable backend with Docker Compose, Kubernetes, or Pulumi on AWS.
• Extensible features: projects, skills, widgets, HTML artifacts, voice mode.
• Supports MCP servers and ACP agents for external tool integration.
• Local-first data storage with optional sync and end-to-end encryption.
• CLI terminal coding agent included.