Tacita vs cloud chatbot apps

Architecture at a glance

Property	Tacita	Typical cloud chatbot app
Where the model runs	On the user's phone (Gemma 4 GGUF via llama.cpp)	On the operator's servers
Where conversations are stored	Encrypted on the device, key derived from the user's password	On the operator's database, typically encrypted under the operator's key
Account required	None	Email or social sign-in, typically
Telemetry / analytics	None — zero analytics SDKs, zero crash reporters	Usually multiple SDKs (analytics, crash, attribution)
Used for training	Impossible — messages never leave the device	Frequently, subject to the operator's policy
Time-to-first-token	Sub-second on flagship phones — no network round-trip	Network-bound; 0.5–3 s typical
Sustained throughput	Bounded by device CPU / memory bandwidth	Bounded by the operator's API quota and queue
Works offline	Yes, after one model download	No
Maximum model size	Bounded by device RAM (Pro tier wants 8 GB)	Bounded by the operator's hardware (often very large)
Pricing	Free tier, plus a one-time €19.99 lifetime Pro purchase	Recurring subscription or per-message billing
Recovery if you forget your password	None — the conversations are gone, by design	Operator-mediated reset
Multi-device sync	None — single device per vault	Yes
Content moderation surface	None — the user is responsible for what they generate	Operator moderates, often aggressively
Character card support (CCv2 / CCv3 / CHARX)	Native, full-fidelity import with lorebook activation	Varies; some support a subset, many do not

What Tacita wins on

• Privacy by construction, not by promise. Cloud apps ask you to trust their privacy policy. Tacita is built so there is nothing to trust us with: your conversations live on your phone, encrypted under a key only you derive.
• Zero recurring cost. Pay €19.99 once for Pro, or pay nothing for the free tier. There is no subscription clock running in the background.
• No queue, no rate limit. The compute is your phone. Use it as much as you like.
• Works on a plane, in a tunnel, on a mountain. After one model download, the radios can be off for the rest of the conversation.
• Native CCv2, CCv3, and CHARX. If you have character card files you already use, Tacita opens them — full lorebook activation, alternate greetings, post-history instructions, and every documented spec field preserved.
• Sealed search. When you do invoke web search, the thumbnails come back into the vault encrypted; tapping a source opens an in-app incognito browser with no shared cookies.

What Tacita gives up — and the trade-offs we accept

• Peak model quality. The largest hosted models outclass any model that fits in a phone's RAM today. If you need the absolute frontier of capability, a hosted model wins. Tacita ships Gemma 4 because it is genuinely good for its weight class.
• Multi-device sync. A vault lives on one phone. Moving to a new device means starting fresh — that is the price of having no server we could sync from.
• Recoverable accounts. Forgetting your password is irreversible in Tacita. We chose this because the alternative is a backdoor we would also have to defend.
• Central content moderation. There is none. The model runs on your phone; what comes out is your responsibility, and the AI disclaimer makes that explicit.
• Always-on model updates. Cloud operators can swap the model under you overnight. Tacita updates the model when you install a new release; the version you have is the version you have.

When Tacita is the right choice

You should probably use Tacita if any of the following matters to you: keeping your conversations off the public internet, paying once instead of every month, having an AI that works offline, or using character cards in their native CCv2 / CCv3 / CHARX format with a proper lorebook engine. You should probably not use Tacita if you need the largest possible model, central moderation, or automatic cross-device sync.

---

More background: how the encryption stack works, how character cards are imported, which models Tacita ships, and the full FAQ.