paperbridge

name	paperbridge
description	Use when a task involves Zotero (search, collections, items, PDFs, full-text), DOI/Crossref resolution, searching the YAMS research workspace or external paper sources, or retrieving locally cached papers from the Paperseed corpus. Provides both a CLI (`paperbridge ...`) and an MCP server (`paperbridge serve`). Prefer MCP tools when available; fall back to CLI invocation otherwise.

paperbridge

Rust CLI + MCP server bridging Zotero (cloud or desktop local API), external paper indexes, and a local Paperseed cache. Use it for literature search, reference resolution, structured paper parsing, and preparing paper content for downstream agents.

MCP availability. When connected to paperbridge serve, this guide is also served as the prompt paperbridge_skill (prompts/get with name: "paperbridge_skill").

When to use

• Search a Zotero library or browse collections, tags, attachments.
• Resolve a DOI to structured metadata (title, authors, year, journal, abstract).
• Search the local YAMS research workspace and external paper indexes: arXiv, Crossref, OpenAlex, Europe PMC, DBLP, OpenReview, PubMed, HuggingFace Papers, Semantic Scholar, CORE, NASA ADS, ScholarAPI.
• Retrieve full-text or structured content from a Zotero attachment or a cached paper.
• Validate, create, update, or delete Zotero items and collections.
• Import or query the local Paperseed corpus (paperseed_enabled = true).

Modes

• MCP (preferred in agent contexts): use the registered paperbridge MCP server tools directly — they mirror the CLI commands below.
• CLI: paperbridge <domain> <action>. Structured commands print readable text by default; add the global --json flag for machine-readable JSON. Runtime errors also use JSON with --json and always stay on stderr, so stdout remains safe for automation.

First-time setup

paperbridge config init --interactive
paperbridge config validate
paperbridge status

Backend modes: cloud (api.zotero.org, needs api_key + user_id), local (Zotero Desktop at http://127.0.0.1:23119, no key), hybrid (local reads, cloud writes).

Core recipes

Search — library, external, and cached

# Zotero library (paginated envelope)
paperbridge library query --query "diffusion models" --limit 10

# External papers (compact by default; --limit is PAGE size)
paperbridge papers search --query "intrusion detection" --per-source 3 --limit 10
paperbridge papers search --query "attention is all you need" --sources arxiv,openalex
paperbridge papers search --query "attention is all you need" --sources paperseed  # cache only
paperbridge papers search --query "What drives detection in GNN" --sources research
paperbridge papers search --query "transformers" --limit 5 --offset 10 --detail full

# Open a hit after search (await-friendly path)
paperbridge papers open --hit-id "arxiv:1706.03762" --want metadata,structure
paperbridge papers open --doi "10.1038/nature12373" --want metadata
paperbridge papers open --url "https://example.org/paper.pdf" --want fulltext

Results are deduplicated by DOI → arXiv ID → PMID → normalized title+first-author. Each hit includes hit_id, ids, match, access, and next suggested tools. Default detail=compact omits full abstracts. diagnostics lists sources that ran, were skipped (missing key), or failed.

MCP tools:

• search_papers { query|q, limit?, limit_per_source?, sources?, cache?, offset?, detail?, abstract_max_chars? } → { query, total_count, offset, limit, has_more, next_offset, detail, hits, diagnostics }
• open_paper { hit_id?|doi?|arxiv_id?|item_key?|paper_id?|attachment_key?|url?, want?, max_chars? }
• search_items / list_collections → paginated envelopes (hits, not bare arrays)

Canonical source wire names: research, arxiv, paperseed, crossref, openalex, europe_pmc, dblp, openreview, pubmed, hugging_face, semantic_scholar, core, ads, scholarapi (aliases still accepted).

Always-on (no key): arXiv, Crossref, OpenAlex, Europe PMC, DBLP, OpenReview, PubMed. Key-gated (appear in diagnostics.sources_skipped when unset): HuggingFace, Semantic Scholar, CORE, NASA ADS, ScholarAPI.

Springer journal articles are discoverable indirectly when indexed by Crossref, OpenAlex, or another configured index. There is no dedicated springer source adapter, and publisher-hosted full text is available only when an open PDF URL, Zotero attachment, or cached Paperseed copy can be resolved.

Resolve a DOI

paperbridge papers resolve-doi --doi "10.1038/nature12373"

When unpaywall_email is configured, the response includes oa_pdf_url.

Read content — prefer open_paper

# After search: open by hit_id / DOI / arXiv / Zotero key
paperbridge papers open --hit-id "arxiv:1706.03762" --want fulltext --max-chars 8000
# If the response includes next_offset, pass that UTF-8 byte offset to continue.
paperbridge papers open --hit-id "arxiv:1706.03762" --want fulltext --max-chars 8000 --offset <next-offset>
paperbridge papers open --hit-id "research:<yams-hash>" --want structure --max-chars 30000
paperbridge papers open --item-key ABCD1234 --want structure
paperbridge papers structure --key ABCD1234

# Vox read-aloud only (not plain fulltext)
paperbridge library read --item-key ABCD1234
paperbridge library read-search -q "sparse attention" --result-index 0 --search-limit 5

Agent spine: search_items / search_papers → open_paper → query_paper / resolve_doi. Use Vox prepare_* tools only for read-aloud.

MCP tools:

• open_paper { hit_id|doi|arxiv_id|item_key|paper_id|attachment_key|url, want, max_chars?, offset? } — preferred; use next_offset to continue fulltext
• get_pdf_text { attachment_key, max_chars?, offset? } / get_item_fulltext { attachment_key, max_chars?, offset? } — low-level bounded attachment/cache paths
• prepare_vox_text / prepare_item_for_vox / prepare_search_result_for_vox — Vox chunks only

Structured paper content

Returns a typed structure with sections, references, and figures. Add --json to CLI calls when the structure will be consumed programmatically. Works with both Zotero items and cached paper IDs.

paperbridge papers structure --key ABCD1234
paperbridge papers query --key ABCD1234 --selector "sections[0].text"
paperbridge papers query --key ABCD1234 --selector "metadata.doi"

MCP tools: get_paper_structure { item_key, attachment_key? }, query_paper { item_key, selector, attachment_key? }. Both accept Zotero keys or cached paper IDs. When a cached paper has no extracted fulltext, metadata is still returned with empty sections (no 404s).

Selectors use dotted paths with bracket indexing (sections[2].text, references[0].title). The source field tells you the provenance: grobid, zotero_fulltext, or grobid_unavailable.

Paper → skill scaffold

Turn a paper's structure into a deterministic SKILL.md scaffold (YAML frontmatter + markdown body). The mapping is mechanical — abstract → "When to use", method/design/implementation → "Method", evaluation/results → "Evaluation", plus limitations and key references. The output is a scaffold for an agent to refine into a real operating procedure, not a finished skill.

paperbridge papers skill --key ABCD1234 > SKILL.md
paperbridge papers skill --key <cached-paper-id>

MCP tool: prepare_paper_for_skill { item_key, attachment_key? } → returns { name, description, markdown }. Accepts Zotero keys or cached paper IDs (same routing as get_paper_structure). Fidelity follows the structure's source: a zotero_fulltext scaffold is heuristic and should be verified against the paper. The CLI prints the raw markdown for piping into a file.

Local Paperseed corpus

Manage the content-addressed local cache and license-gated seed manifests:

paperbridge paperseed corpus status
paperbridge paperseed corpus list
paperbridge paperseed corpus show <id-or-unique-hash-prefix>
paperbridge paperseed corpus import ./paper.pdf --license cc-by
paperbridge paperseed corpus import ./large.pdf --license cc-by --no-fulltext
paperbridge paperseed corpus ingest --metadata item.json --file paper.pdf --license cc-by
paperbridge paperseed corpus query -q "induction heads"
paperbridge paperseed corpus export --format bibtex
paperbridge paperseed corpus remove <id-or-unique-hash-prefix>
paperbridge paperseed corpus reindex

paperbridge paperseed seed check --paper-id <id>
paperbridge paperseed seed create --paper-id <id>

paperseed corpus export defaults to BibTeX. Pass the global --json flag to export the corpus as JSON; --format json is accepted only with --json.

Imported PDFs have their text automatically extracted into content-addressed text blobs for full-text search. Use --no-fulltext to defer extraction until first read; PDF extraction does not perform OCR. corpus status reports paper and index document counts and warns when they drift. When paperseed_yams_enabled = true, imports and OA mirrors are synchronously added to YAMS with authoritative title, DOI, authors, venue, and source metadata. The returned hash is persisted as yams_hash; search cache summaries expose yams_indexed.

The research source searches global YAMS documents under research paper projects independently of the Paperseed database. Related PDF/TeX fragments are collapsed into one paper hit with a stable research:<hash> ID. Opening that hit assembles readable abstract, introduction, design, evaluation/results, related-work, conclusion, and appendix blobs when available. Inspect access.content_state: ready is openable; stale means YAMS retained search metadata but neither its content blob nor the original file is readable.

OA auto-mirroring & DOI resolution. With paperseed_auto_download = true, search hits are mirrored into the corpus. Hits that already carry an open-access PDF url are downloaded directly; hits that only expose a DOI (common for metadata-only sources like Crossref, PubMed, and DBLP) are resolved to an open PDF via Unpaywall, falling back to OpenAlex's best OA location. Set unpaywall_email with a real contact address for Unpaywall. Resolver-derived licenses are preserved for mirrored files; hits with a direct PDF but no license evidence remain unknown and are not seedable.

Write Zotero items & collections

Write ops take a JSON file on disk. Cloud backend requires api_key with write scope.

paperbridge item validate --file item.json --online
paperbridge item create --file item.json
paperbridge item update --file item.json
paperbridge item delete --file item.json
paperbridge collection create --name "ML 2025"

Run as MCP server

paperbridge serve
paperbridge config snippet --target claude
paperbridge config snippet --target opencode

Key config keys

key	purpose
backend_mode	cloud, local, hybrid
api_key	Zotero API key — redacted in config get unless --show-secret
user_id	numeric Zotero user ID
group_id	numeric group ID (optional)
library_type	user or group
paperseed_enabled	enable local Paperseed corpus (default false)
paperseed_auto_download	automatically mirror OA PDFs into local corpus (default true)
paperseed_yams_enabled	use YAMS as experimental storage/search backend (default true)
paperseed_corpus_root	override corpus path
hf_token, semantic_scholar_api_key, core_api_key, ads_api_token, scholarapi_key	gate external sources
ncbi_api_key	optional PubMed rate-limit upgrade
unpaywall_email	enables OA-PDF enrichment
grobid_url	GROBID endpoint; if set, auto-spawn is disabled
grobid_auto_spawn	launch GROBID via Docker (default false)
grobid_image	Docker image for auto-spawn
log_level	error, warn, info, debug, trace

paperbridge config get masks secrets by default. Pass --show-secret to reveal.

Gotchas

• CLI output is human-readable by default — pass the global --json flag for stable structured output (for example, paperbridge --json papers search --query "transformers"). Runtime failures use { error, reason, try } JSON on stderr with a non-zero exit code. MCP tool results remain JSON.

• Cloud api_base must be HTTPS (or http://localhost for local mode).

• --limit is page size (default 10). Fan-out uses --per-source / MCP limit_per_source.

• Search is compact by default — pass --detail full / detail: "full" for abstracts. Full abstracts default to 280 characters; set abstract_max_chars=0 for unlimited text.

• Search results are paginated — use offset/limit; check has_more / next_offset. Later pages expand the per-source fetch window automatically, up to a safe window of 200.

• Key-gated sources skip loudly — see diagnostics.sources_skipped / sources_failed.

• Cached papers are conservative by default: default cache-only hits need strong relevance, and --sources without paperseed excludes cache hits. Use --sources paperseed for explicit cache-only search.

• Research hits report availability: use --sources research for YAMS-only discovery. Do not attempt to open hits with access.content_state: stale until the source is re-indexed.

• PDF text extraction happens automatically during local corpus import unless --no-fulltext is passed. Deferred extraction runs on first read; neither path performs OCR.

• PDF validation is strict when extracting — HTML/error pages saved as .pdf, malformed or encrypted PDFs, and files over 100 MiB are rejected with recovery guidance. Use --no-fulltext only when you deliberately need to retain a PDF for later handling.

• — , , and default to 8,000 characters. When is present, repeat the same call with that UTF-8 byte . Vox tools use .

Verify install

paperbridge --version
paperbridge status
paperbridge config validate

Contributors

CLI surface changes must be reviewed against docs/design/cli-design.md. MCP/search/return-shape and agent discover→read changes must also follow docs/design/llm-interface.md and the backlog in docs/design/llm-interface-tasks.md.