Paperclip (GXL) — Agent-Native Scientific Literature MCP¶

What it is¶

Paperclip is an MCP server (and CLI) for scientific literature, built by GXL (Generative Expert Labs), a Stanford-adjacent group connected to James Zou's lab. Unlike PubMed or Google Scholar — which return links and stop — Paperclip exposes papers to an LLM as a structured filesystem: search, grep, cat, map, and SQL operations chain through a stateful results_id, so an agent can narrow a corpus iteratively over many turns without re-searching.

Naming caveat: there is an unrelated product (paperclip.ing / paperclipai) for AI agent orchestration. Not the same tool. The literature one is paperclip.gxl.ai, made by GXL.

Corpus¶

Source	Coverage	Type
PubMed Central	~5M+	Full text, biomedical, open access
arXiv	~3M	Full text, ML / math / quant-bio / physics / CS
bioRxiv + medRxiv	~3M+	Full text, preprints
OpenAlex	~150M+	Abstracts + structured metadata only

Total: ~11M full-text papers + ~150M abstracts. Coverage caveats: PMC has embargo periods for some journals; only open-access full text is indexed. Abstract-only coverage means you can map the landscape but not grep within the body for non-OA papers.

Commands¶

Command	Function
`search`	Hybrid (BM25 + vector embedding) search; returns 1–2 sentence TL;DRs
`grep`	Regex / keyword search within full text. Vendor claim: 36–294× faster than raw grep on the 8M+ paper index — unverified
`cat`	Reads structured paper text (sections, tables, figures)
`map`	Applies a prompt across a result set — the synthesis primitive
`ask-image`	Multimodal query against figures and images
`sql`	Read-only SQL over metadata table; 15s timeout, 200-row limit
`--from <results_id>`	Chains operations against a previous result set

The --from chaining is the load-bearing piece — every search produces a cloud-stored results_id, and subsequent grep / map / cat operations can target it. An agent can search broadly, narrow by grep, synthesize via map, then drill into specific papers without losing the working set.

Setup¶

claude mcp add --transport http paperclip https://paperclip.gxl.ai/mcp

Or in claude_desktop_config.json:

{
  "mcpServers": {
    "paperclip": { "url": "https://paperclip.gxl.ai/mcp" }
  }
}

The hosted MCP at paperclip.gxl.ai/mcp is currently free and does not appear to require an API key. Pricing tiers have not been announced. As a Stanford / Together Compute-adjacent project this may remain free for research use, but plan for the possibility of usage limits.

Comparison to PubMed and Google Scholar¶

Dimension	PubMed	Google Scholar	Paperclip
Primary user	Humans (web UI)	Humans (web UI)	LLM agents (and humans via CLI)
Corpus	~36M citations, biomedical	Broad multidisciplinary	11M full-text + 150M abstracts, bio + ML + preprints
Searches within full text	No (links to it)	No (links to it)	Yes
Stateful chaining	No	No	Yes (`--from`)
Native API for agents	Limited (NCBI E-utilities)	None official	MCP-native
Synthesis primitive	Manual	Manual	`map`
Multimodal (figures)	No	No	Yes (`ask-image`)

When to still use PubMed: MeSH-vocabulary searches, citation tracking by PMID, peer-reviewed-only filtering. (Paperclip does not appear to support MeSH; preprints and OA full text are mixed in by default.)

Where Paperclip is differentiated: searching within papers, cross-domain queries that span biomedical + ML, multi-paper synthesis via map, and any agent-driven workflow.

The OE bio-ai-tools.md page already lists the Anthropic life-sciences MCP marketplace (PubMed, bioRxiv, ChEMBL, Open Targets, ClinicalTrials.gov) as Phase 0 core. Paperclip is complementary, not a replacement: the marketplace plugins are best-in-class for their specific sources (PubMed's MeSH index, ChEMBL's bioactivity tables); Paperclip wins on full-text search and cross-source synthesis.

Reliability — 2026-05-05 verification test¶

A first end-to-end test (uricase variant landscape + A. oryzae expression evidence) surfaced a systematic hallucination pattern in the map operator that significantly changes the trust model for this MCP. Documenting here so future Paperclip sessions inherit the correct guardrails.

Trust ranking by tool¶

Tool	Reliability	Notes
`search`	High	Returns real PMC / bioRxiv / arXiv records with accurate IDs and titles. Verified by spot-checking against `meta.json` and external lookups.
`cat /papers/<id>/meta.json`	High	Authoritative paper metadata — title, abstract, authors, journal, PMID, DOI. Use as ground truth for abstract-level claims.
`grep PATTERN /papers/<id>/...`	High	Returns real text from indexed paper bodies. Use to verify any quantitative claim before propagation.
`cat /papers/<id>/content.lines`	High	Real full-text. Same trust level as grep.
`map --from <id> "extract X"`	LOW — hallucinates quantitative data and misattributes organisms	Lighter "reader" model behind `map` substitutes plausible-looking domain values when full text doesn't directly support the requested field. Treat outputs as hypothesis-generation, not evidence.
`reduce --from <map-id> ...`	Compounding risk (model-on-model)	If the underlying `map` is wrong, `reduce` consolidates wrong claims into a confident-looking summary.

Concrete examples from the 2026-05-05 test¶

All of these were caught by grep-verifying body text or reading the actual abstract via meta.json after map returned the structured field. None of these are subtle — they're load-bearing identity errors.

Paper	Abstract / body says	`map` returned
PMC9773812 (Najjari 2022, PASylated UOX)	A. flavus UOX, K_m 52.61 µM	*A. globiformis* uricase variant (S284G, K304R), K_m 0.007 mM (~7.5× off — see §"2026-05-13 correction" below)
PMC4881585 (Xie 2016, chimeric uricase)	Porcine-human exon-replacement chimera	*P. chrysogenum*-human exon chimera (different organism entirely)
PMC10561068 (Yan 2023, Arthrobacter CSAJ-16)	Optimal T 20°C, K_m 0.048 mM (Lineweaver-Burk, body L40)	Optimal T 40°C, K_m 0.015 mM
PMC12106716 (Rahbar 2025, A. flavus disulfide design)	Pure computational paper — frustration mapping + RMSF + tunnel analysis, no wet-lab	Invented Tm 64.9 → 70.3°C, K_m/k_cat measurements as if wet-lab data existed; named non-existent S173C/L221C mutation pair (real predicted pairs are A6-C290 and S119-C220)

These are not transcription errors. They are confabulations — plausible-looking values and organism names that would pass a casual review but are not in the underlying full text.

2026-05-13 correction — Km magnitude¶

The original table entry for PMC9773812 (Najjari 2022) recorded the Paperclip map operator's misreport as K_m 0.007 mM (~7,500× off). This was itself an arithmetic / unit-confusion error in the documentation: the true Km is 52.61 µM = 0.05261 mM, and the misreported value is 0.007 mM = 7 µM. The actual factor between them is 0.05261/0.007 ≈ 7.5×, not 7,500×.

The original "7,500×" figure propagated from this wiki page into the cross-vendor heterogeneity-guard paper draft (papers/cross-vendor-heterogeneity-guard/) §5.3 and was surfaced during a cross-vendor review of that paper on 2026-05-13 — DeepSeek V4-Pro (review of §4+§5) flagged the arithmetic discrepancy as Rejected, with the correct factor and a unit-conversion note. The wiki table has been corrected above; the paper §5.3 has been corrected and the catch is logged in the paper's revisions.md as a reflexive demonstration of the methodology working on its own production.

The substantive point of the case study is unchanged: a ~7.5× misreport of a kinetic parameter is still a disqualifying reliability failure for an automated literature-extraction tool. The case-study lesson holds — only the magnitude was overstated. The correction here is itself the canonical example of why the pre-commit grep-verify gate exists: the original "7,500×" claim was never verified against the arithmetic, and the error rode the corpus for nine days (2026-05-05 → 2026-05-13) before an external cross-vendor pass caught it.

Probable mechanism¶

Paperclip's map operator runs a lightweight per-paper extraction model. When asked for specific quantitative or identity fields the model can't ground in the indexed text, it appears to substitute domain-plausible values rather than emit "not reported." This is a known failure mode for small models forced into structured-output tasks they can't actually support.

Verification discipline for any future Paperclip session¶

Use search and grep as primary evidence. Treat map outputs as a sketch of where to look, not as data.
Grep-verify every load-bearing number against the paper body before letting it into the wiki. If grep can't find it, it didn't come from the paper.
Anchor identity claims (organism, gene, host, year) to meta.json. Abstracts in meta.json are clean — use them as the source of truth for organism / engineering-approach claims.
Never propagate reduce summaries directly. If the underlying map was bad, reduce will confidently consolidate the bad claims.
Computational vs. wet-lab status must be checked from the abstract or methods section. map does not reliably distinguish these.

The corpus and the deterministic tools are good. The model-mediated synthesis layer is not yet trustworthy for content destined for the wiki.

Relevance to Open Enzyme¶

Three plausible roles, listed by how concrete the integration is:

1. Manual research depth — immediately usable¶

For any specific question already in the wiki, Paperclip enables systematic full-text review where the existing MCP plugins can only return abstracts:

Uricase mutation landscape — search + grep + map across PMC and arXiv to catalog published variants, hosts, catalytic parameters, immunogenicity. Currently scattered across uricase.md, crispr-uricase.md, engineered-koji-protocol.md.
NLRP3 inhibition mechanisms — map across the NLRP3 corpus for dosing and efficacy data, complementing nlrp3-exploit-map.md and nlrp3-inhibitor-screen.md.
Koji / A. oryzae expression systems — cross-reference food-science and synthetic-biology literature for promoter / cassette / yield data; expands aspergillus-oryzae.md and the Ward 1995 dual-cassette material in koji-endgame-strain.md.
Cross-domain queries — e.g., grep for "uricase" + "food-grade" or "ABCG2" + "probiotic" across the full corpus to surface intersections that PubMed and Scholar fragment.

2. Sweep-daemon integration — DECIDED 2026-05-05: do not integrate¶

Decision: do not wire Paperclip into the three-pass sweep daemon. The reliability finding above (map operator hallucinations) is disqualifying for any architecture that lands Paperclip-derived content into the wiki without a human in the loop.

The original framing — Paperclip-augmented pass producing a "literature delta" surface for the synthesis stage — assumed the synthesis primitive (map) could be trusted to faithfully extract per-paper findings. The 2026-05-05 verification test invalidated that assumption: map produced load-bearing identity errors (wrong organisms, wrong gene names, fabricated kinetic numbers) on multiple papers in a single session. Wiring this into the sweep would inject a structured external hallucination source into a corpus designed for PhD-grade rigor, exactly the failure mode the multi-model synthesis architecture in open-source-platform.md is meant to guard against.

The reopen condition: if GXL ships a verified upgrade of the map reader model and we re-run the verification test (uricase variant landscape is a clean repeatable probe; ~12 papers, multiple known-correct ground truths via abstract + grep) and it passes cleanly, revisit. Until then Paperclip remains a manual-research-only tool, used interactively with verification discipline, never embedded in an automated pipeline.

The original "open platform decision" entry in synthesis/ (architecture: synthesis/README.md) should be closed as resolved on the same date with this same outcome.

3. Protein-engineering support — via arXiv coverage¶

arXiv inclusion brings the ML / computational biology literature into the same index as biomedical full text. Useful for keeping bio-ai-tools.md current — e.g., new protein language models, directed-evolution algorithms, or kidney-tropic siRNA delivery (the modality-chokepoint-matrix.md URAT1 vector) without bouncing between PubMed and arXiv.

Recommendations¶

Updated 2026-05-05 after the verification test:

Install the MCP. Done. Available across the abent umbrella, not just OE.
Use Paperclip as a manual research tool only, with the verification discipline above. Never propagate map or reduce outputs into the wiki unverified — anchor every quantitative claim in a grep or abstract meta.json cross-check.
Do not integrate Paperclip into the sweep daemon. See §"Sweep-daemon integration — DECIDED" above for rationale and reopen condition.
Report the map hallucination pattern to GXL. Bug report drafted 2026-05-05 (see session log); contains reproducible examples with paper IDs and what abstract says vs. what map returned.

Watch list¶

Pricing changes — if GXL adds paid tiers, evaluate whether sweep-driven query volume justifies the cost.
Public repo — github.com/GXL-ai/paperclip (per James Zou's announcement) for feature requests, especially around map improvements and citation-graph traversal.
PMC embargo lag — Paperclip can only index what's openly available; recent high-impact papers in non-OA journals will be abstract-only.
Rate limits — untested for sustained automated query volume.