← All projects

06 · Public source

Multipass

A desktop clipboard and chat workspace that sanitizes text before it reaches a model.

Clipboard and freeform input are treated as untrusted, sanitized, explicitly wrapped, and only then passed into prompt construction.

View source

Problem

Model clients were receiving raw user-controlled clipboard text directly, with no boundary between untrusted input and prompt construction.

Mechanism

  1. Route input through a security layer that sanitizes, marks trust boundaries, and wraps it before prompt building.
  2. Run hosted API backends (DeepSeek, Kimi, Bedrock) inside the app chat UI; launch CLI backends (claude, codex) in a real terminal window instead of faking an embedded session.
  3. Apply one-click clipboard transforms such as whitespace normalization, hostname extraction, and base64 or URL encode/decode.

Evidence

Status
Early stage · Usable
Backends live
3 hosted APIs · 2 coding CLIs
Architecture
Domain · security · application split

Boundary

The PyPI name `multipass` is taken by another project, so no PyPI distribution exists; the source is available on GitHub only.

The backend set is small and the desktop UI targets one platform family; hosted APIs require environment credentials.