Capture excerpts, screenshots, and files. Keep their sources close. Review what you carry into the next conversation.
The problem
Moving between AI conversations often means rebuilding context: finding the useful excerpt, explaining an earlier decision, and remembering which questions remain open. Context Shelf brings that material into books with working context and original sources side by side.
The design question is what a person needs to inspect before trusting the next handoff.
One collection, three interfaces
The native Mac app, command-line utility, and MCP server use a shared Swift core for models, persistence, routing, and export. People and agents work with the same collection through interfaces suited to their tasks.
Storage is local. Basic capture and extractive context generation work without a remote model or API key. Optional Apple Foundation Models support depends on the Mac and system configuration.
Keep decisions open to review
Source clips remain available alongside working context. A handoff preview lets the user tailor the outgoing text before copying or exporting it. Portable exports include Markdown, structured data, and original attachments.
- Keep provenance with the evidence; label unknown origins honestly.
- Let an explicit destination override automatic routing.
- Leave ambiguous matches unfiled so they can be reviewed.
Where judgment still matters
Routing uses a deterministic vocabulary and source-host heuristic. It can miss related material that uses different words; its scores are not probabilities. Generated context can also omit details. The original clips and editable handoff give the user a way to check both.
The field guide documents the architecture, agent interface, and current verification limits, including capture flows that still need end-to-end checks with macOS permissions enabled.