Hi Cryptpad Team

As mentioned already in conversations, we'd like to describe here once more our project idea: we would like to develop an optional, fully client-side "Copilot" module for CryptPad as a working prototype in a forked repository. The goal is to add lightweight text-assistance features - entirely on-device, without changing the existing security or encryption model.

We’d love to get technical feedback and guidance from the CryptPad core team.

Motivation

Modern “cloud copilots” are powerful but not usable in E2E-encrypted, privacy-focused platforms like CryptPad. At the same time, WebGPU inference for small language models (LLMs) has become fast enough to be practical on end-user devices.

Our goal: developing a usable prototype for summarization and rewriting directly into CryptPad’s document app - completely offline, no data transfer (besides the initial loading of the LLM itself).

This would make AI assistance available to NGOs, educators, public institutions and all other users that cannot use cloud-based AI tools for compliance reasons or data privacy concerns.

Core idea & architecture

High Level Target Architecture Diagram
Image description

Data/Controll Flow Diagram
Image description

Description of implementation approach

Inference: runs in a same-origin Dedicated Web Worker (WebGPU accelerated)
Models: small, instruction-tuned decoder LLMs (e.g. Phi-4 Mini or TinyLlama-1.1B)
Model assets: self-hosted by the instance (connect-src 'self'), cached in OPFS/IndexedDB, progressive download with resume
UI integration: editor panel inside Inner (Docs), communication via Inner -> Outer -> Worker -> Outer -> Inner message flow
Security: strict CSP; no third‑party/cross‑origin requests; model assets served from same origin; no telemetry; prompts/responses in‑memory
Feature flag: disabled by default; optional module that admins can enable locally (zero-impact when off)
Undo-safe insertion: via existing editor API (not direct DOM manipulation)

Questions

What we’d love from the CryptPad team are your comments/thoughts regarding:

Do you see any potential conflicts with CryptPad’s CSP or E2E model which might not be easy to spot? To our understanding there shouldn’t be any since the Copilot panel would live inside the Inner iframe (Docs UI), all LLM requests are brokered by the Outer to a same-origin Dedicated Worker. Model assets would be self-hosted and cached in OPFS. Prompts/responses would remain in memory and no plaintext crosses beyond the established Outer/Inner trust boundary.
Are there any other technical aspects that could make this approach impractical?

Thanks for taking the time to review and share your thoughts.