What it does Thunderbolt is a desktop and mobile AI client from the Thunderbird team (yes, the email people). It lets you chat with LLMs through a single interface, self-hosted via Docker or Kubernetes. The pitch is vendor independence: bring your own API keys for OpenAI-compatible providers, or run local inference through Ollama or llama.cpp.

The interesting bit The “offline-first” ambition is real but currently aspirational. Right now you still need their backend running for authentication and search—though you can disable search in settings. It’s a frank admission in a space where competitors pretend self-hosting is already solved.

Key highlights

• Cross-platform coverage is thorough: web, iOS, Android, Mac, Linux, Windows (via Tauri)
• make doctor checks your dev environment and tells you exactly what’s missing—small thing, but it signals someone actually set this up more than once
• MPL 2.0 license, not some faux-open corporate license
• Currently undergoing a security audit before enterprise push
• Enterprise support and FDEs (Full Disk Encryption?) mentioned but not detailed

Caveats

• No public inference endpoint; you must supply your own models
• Authentication and search still require internet connectivity to their backend
• Documentation and community infrastructure are explicitly “actively being worked on”
• Security audit is in progress, not complete

Verdict Worth watching if your org’s AI strategy includes the words “on-prem” and “compliance.” Skip it for now if you want a drop-in ChatGPT replacement that just works out of the box.