ThunderboltThe unfinished thought. The ambitious plan. The question you haven’t asked anyone yet. You deserve room to explore without giving your privacy away.
Find your path to private AIKeep the computing under your roof, or choose cloud AI with protection you can verify.
Self-hosting + local models
Run Thunderbolt on infrastructure you control and connect a model running on your device or your own servers. Your prompts can stay inside your environment.
You decide what runs, where it runs, and who has access. Keep model inference within your environment, and choose which external services to connect.
Build your own setupVerifiable confidential computing
Choose Thunderbolt’s private models powered by Tinfoil. They run inside hardware-protected enclaves, designed to keep your prompts and responses unreadable to Thunderbird and Tinfoil.
Thunderbolt verifies the enclave before sending an encrypted request. The model reads it inside that protected space; our relay carries encrypted data.
Meet our private models
We partner with Tinfoil because privacy should be built into how AI runs. Self-hosting teams can also connect their own Tinfoil account—these paths work together, too.
Thunderbolt doesn’t use your conversations to train AI models. Tinfoil-powered private models don’t use your prompts or responses for training, either.
If you connect a different model provider, its own retention and training policies apply. Choosing your provider is part of choosing your privacy.
Where things live. What gets shared. What you can change. Here’s how Thunderbolt approaches each.
Read the technical detailsConversations are stored locally first. Uploaded files and provider API keys aren’t included in cross-device sync. An attachment can still be sent to the model you ask to work with it.
When you enable sync, sensitive content—including messages and chat titles—is encrypted on your device before upload. Some metadata, such as timestamps and record relationships, stays readable on the server.
Encrypted sync is a preview feature and hasn’t had a cryptography audit. If your organization enables key escrow, it can decrypt your synced content. See what encryption covers.
A search service receives your query. A connected tool receives the arguments it needs. An external model or agent receives the conversation it answers. These connections aren’t automatically protected by Tinfoil’s enclave.
Choose the services you trust, and turn off connections you don’t need. Self-hosting alone doesn’t keep requests to external services local. Manage your connections.
Private content doesn’t mean zero operational data. Accounts, sign-in sessions, and managed-model usage records help a deployment function. App analytics exclude conversation content. They are enabled by default when your deployment configures analytics, and can be turned off in Settings → Preferences.
This website uses Matomo for page-view and link analytics. For the specifics, read our data-flow documentation and privacy policy.
Protecting people and protecting their private conversations should move forward together. We believe safety should be transparent, accountable, and built with respect for the person using the technology.
Our partner Tinfoil shows one way forward: running safeguard models inside trusted execution environments, or TEEs. In Tinfoil Chat, the checks happen within the protected boundary, without exposing conversation content to operators.
This is how Tinfoil Chat handles safety today. Thunderbolt’s confidential inference connection does not currently include those safeguards.
Safety Without Compromising on PrivacyInside Tinfoil Chat’s secure enclave
Private conversation
A limited signal can leave.
The conversation stays private.
Tinfoil Chat may share a violation flag and conversation ID associated with an account, but not the conversation content or violation category.
Open source means you can look closer. Read the code, understand the trade-offs, and choose the setup that feels right for you.