Archived asset
An image copied into SourceShelf-managed local storage from a supported capture or conversion. Exports include only assets referenced by selected readable Markdown.
Baseline
The hashes, metadata, dates, format, source IDs, and ordering saved after a successful pack delivery. Refresh & Compare uses it as the “then” state.
Capture recipe
A reusable Safari-capture policy for destination templates, YAML, content mode, image/link behavior, review, domain rules, and staleness.
Library
The searchable browser for converted and captured sources. It is not the current pack.
Living pack
A saved pack with a successful-export baseline that can classify local sources as new, changed, missing, unchanged, or removed.
Local AI Access
SourceShelf’s opt-in MCP sharing workflow for a saved pack.
MCP
Model Context Protocol, a standard through which an AI host can discover resources and tools from a server. SourceShelf uses local stdio only.
OKF
Open Knowledge Format. SourceShelf exports canonical v0.2 concept metadata with SourceShelf manifest extensions.
Pack
An ordered selection of Library source IDs. A saved pack can be exported, compared, or shared.
Recipe review
An optional queued step that lets you edit captured Markdown and destination metadata before saving.
Retrieval chunk
A bounded, structure-aware section of a source recorded in chunks.jsonl with stable IDs, provenance, hashes, headings, and estimated token count. It is not an embedding.
Semantic document
SourceShelf’s versioned local representation of headings, paragraphs, lists, tables, code, images, captions, page boundaries, quotes, and fallback content.
Share ID
A random local authorization identifying one published MCP snapshot.
Snapshot
An immutable, pack-scoped MCP copy of readable Markdown, referenced assets, catalog, llms.txt, checksums, and allowlist.
Trust & Safety
The preflight report for readability, packaging, freshness, checksums, assets, and conservative untrusted-instruction patterns. It is advisory, not sanitization.
Untrusted reference
The trust classification applied to captured and converted material to remind AI clients and users not to treat source-body instructions as system authority.