Cloak Dagger AI

Intelligence that lives where your conversations live.

A local-first assistant inside the messenger: private memory, precise retrieval, on-device reasoning, and labeled processing — with cloud only as a choice you make.

Processed on-device
Cloud off by default
Cloak Dagger AI · WebGPU

Pipeline

Cloak Dagger AI is not your memory.

Four layers with separate responsibilities keep the assistant precise and private — the model reasons, it does not rummage. It is deliberately not a chatbot bolted onto a messenger.

01

Memory — local store

Facts, people, events, and summaries you choose to keep. Browser-local by design and structured for encryption at rest. Ghost Chats are excluded unless you allow otherwise.

02

Retrieval — scoped and small

Structured indexes, keyword search, recency, conversation and contact scope, plus local semantic search. The pipeline selects a handful of permissioned items — never your whole history.

03

Reasoning — Cloak Dagger AI on-device

Cloak Dagger AI runs in a Web Worker over WebGPU. Synthesis happens only when a request needs it, and the model never searches memory itself.

04

Orchestration — honest routing

Every request is classified: answer deterministically, retrieve only, reason locally, translate, or — only with explicit consent — fall back to cloud. Each answer carries its route as a label.

Cloud boundary

Local-first — not local-only.

Local processing is the default. Cloud fallback can be disabled completely and must never occur silently.

Local-only

Processing never leaves the local path. Cloud fallback is fully disabled.

Ask before cloud processing

Cloak Dagger requests your consent first, and labels the answer with the route it used.

Allow cloud

Cloud processing is permitted — still labeled, still revocable at any time.

Questions

Detail, without the dark patterns.

Does the model see my whole memory database?

No. The model receives only the small retrieved context selected by the retrieval pipeline under your permission scope. It cannot browse or query memory directly.

What if my device cannot run local inference?

Messaging works normally, and Cloak Dagger AI states precisely why local inference is unavailable — WebGPU, storage, or installation. Cloud is only ever an explicit choice.

How do I know where an answer was processed?

Every AI answer carries a processing badge: provider, model, location, whether cloud was used, and how many local memories were retrieved. The badge opens the details.

Is translation also local?

Where a local translation model is installed, yes. Translation is a separate provider with its own artifact — independent from generation.

For conversations that should remain under your control.