The clipboard is the most-used primitive on a desktop and one of the least changed — a one-slot buffer with amnesia. No history, no search, no recall. You copy, you paste, and whatever was there a moment ago is gone.
ClipMind was built to answer a simple question: what if the clipboard remembered — and could think about what it holds? The premise is that recall over everything you have ever copied is worth having, and that AI can act on those fragments — code, images, notes — but only if it stays local and under the user's control.
Clipboard managers stop at searchable lists. The hard constraints lie elsewhere: instant recall must work inside any application without disrupting workflow, because if recall breaks flow, users will never use it. History must survive across machines without synchronizing to the cloud. AI transforms — translation, structured extraction, image Q&A — must work without holding user data on remote servers or requiring users to manage credentials. Installation must require zero native build steps, zero compiled bindings, zero fragile toolchains that fail across platforms.
Each is solvable in isolation. Together, they define the actual scope of the problem.
ClipMind is an Electron + React + TypeScript desktop application. The clipboard watcher runs continuously, feeding text and images into a sql.js store — a pure WebAssembly SQLite implementation that requires no compilation. The WASM database persists atomically to disk with debounced writes; images are stored alongside it in the filesystem. Supabase provides OAuth accounts and per-user, row-level-security database views, so history synchronizes across machines without exposing raw data to the sync layer.
A global hotkey opens a command palette rendered under the cursor. The palette combines fuzzy string search with semantic search powered by Gemini embeddings to find any clip in history. Once selected, the clip pastes directly back into the previous application — the user never leaves the app they were using. When the user asks the AI to act on a clip — translating, extracting structured data, or answering questions about an image — the request is sent through an Xandrov-hosted proxy to Vertex AI. Users never handle Gemini credentials; the proxy enforces request scope per clip to prevent data leakage.
The stack is minimal by design: Electron for native hotkey and clipboard access, React for UI, sql.js for the local store, Supabase for sync and auth, and Vertex AI for inference. No Python backends, no complex DevOps — the app is the full system.
Recall is worthless if it breaks flow.
The system exists as a working desktop build with complete source code available for inspection. The clipboard watcher, WASM-based sql.js store with atomic disk persistence, per-machine sync layer powered by Supabase, command palette combining fuzzy and semantic search, and AI transforms through Vertex are all implemented and operational. The build is in active use and ready for deployment to new users.
Nothing here exposes a user's clipboard history or personal data — the local store stays on the user's own machine by design. Credentials and API keys are kept out of anything published, and the AI proxy is described by what it does rather than by its keys. The engineering is the point; the contents of anyone's clipboard are not.