Jamie Pine has spent years building Spacedrive, a cross-platform file manager, and watching it not ship. Two versions in — both grounded by infrastructure debt and the punishing surface area of five operating systems — he's ditching the file browser entirely.
What's left is Spacedrive v3, launching March 15: a local-first data engine that ingests whatever you point it at — Gmail, Slack, Obsidian notes, GitHub issues, Apple Notes, calendar events, contacts, browser history — indexes everything on-device, and makes it searchable without sending a byte to the cloud. Pine's argument is that the index was always the product. The file browser was just the first thing he tried to wrap around it.
The technical centerpiece is a four-stage processing pipeline that every record passes before it can be searched. Stage one is the unusual one: it runs Meta's open-weight Prompt Guard 2 classifier on-device, in under 50 milliseconds, screening for adversarial prompt injection before anything enters the index. The threat model is straightforward — anyone who can send you an email can embed arbitrary instructions into your AI agent's context window. Spacedrive quarantines flagged content rather than surfacing it, addressing what OWASP ranks as the top LLM vulnerability at ingest time rather than runtime. Stages two through four handle content classification and quality scoring, adapter-specific cleanup such as stripping email signatures and filtering Slack bots, and final indexing into both an FTS5 full-text index and a LanceDB vector index.
Prompt injection defense at the ingestion layer is not a common feature in local-data tooling — most pass raw content directly to LLM context with minimal preprocessing — though independent verification of the competitive landscape is difficult.
At launch, Spacedrive ships with 10 adapters: Gmail, Slack, Obsidian, GitHub, Apple Notes, Chrome, Safari, macOS Contacts, macOS Calendar, and OpenCode. A trust tier system — authored, collaborative, external — governs screening strictness per repository, so a personal notes vault and a cold email thread aren't processed the same way.
The stack is Rust throughout, with Tauri 2 for the desktop shell, FastEmbed for on-device embeddings, and LanceDB for vector search — a single binary, all data stored locally under ~/.spacedrive/. The project integrates directly with Spacebot, Pine's AI agent product for teams: Spacebot links spacedrive-core as a Rust crate, meaning data reaching the agent has already been screened, classified, and scored before it gets anywhere near a prompt. The core is free and open source under AGPL-3.0. The commercial angle is Spacebot, which gets considerably more defensible with a properly curated knowledge layer underneath it.