Most literature tools were designed around a human sitting in a browser. You search, open abstracts, download PDFs, read, take notes, and slowly assemble an answer. That workflow still works. It is also a poor fit for an AI agent that can search, read, compare, and extract across a defined source set if it has the right kind of access.
Paperclip by GXL is built around that idea. It exposes scientific literature and adjacent research sources as a virtual filesystem, with a command-line interface, Python SDK, and MCP support for agents. Rather than treating a paper as an opaque link or a blob of PDF text, it makes papers available as structured objects with metadata, full text, sections, figures, and supplements.
That sounds technical, but the practical consequence is straightforward: an agent can move from “find me papers” to “show me what these five papers actually did, where they disagree, and what I still need to read myself.”
What Paperclip is, and what it is not
Paperclip is not another chat interface that gives you a polished paragraph with citations. It is research infrastructure for people and agents that need to interrogate the underlying material.
The current platform covers full text from PubMed Central, bioRxiv, medRxiv, and arXiv, plus OpenAlex abstracts, FDA documents, ClinicalTrials.gov and international trial registries, and biological databases including UniProt, PDB, and ChEMBL. GXL describes the platform as accessible through CLI, SDK, and MCP, which matters because each route suits a different working style.
It also is not a substitute for methodological judgment. A fast answer across twenty papers may be helpful, but it does not decide whether the studies are comparable, whether a surrogate endpoint is meaningful, or whether a paper's limitations matter for your patient population or research question. The value is that Paperclip can get you to that judgment faster with the source material visible.
The workflow it enables
The important unit is not the paper. It is the research question.
Say you are preparing a grant section on a drug target, checking whether a trial landscape has changed, or trying to understand why two groups reported conflicting results. The old workflow is a sequence of browser tabs. The better workflow starts by defining the answer you need, then making the retrieval and extraction steps explicit.
Define the question before searching
Ask for the decision-relevant output: mechanisms of resistance, completed randomized trials, the assay used, the population studied, or the reasons two results differ. Do not begin with a vague request to “summarize the literature.”
Search broadly, then inspect the actual sources
Paperclip supports natural-language search and corpus-wide regex. Search gives a shortlist; full text, sections, figures, and supplements let you see what the paper actually reports.
Extract the same fields across a small set
Its map workflow asks the same focused question across multiple papers in parallel. For example: What model system was used? What was the primary endpoint? What resistance mechanism was reported? The point is comparable extraction, not mass summarization.
Return to the papers where the answer matters
Use the extracted comparison to identify disagreements, missing methods, or papers that need close reading. The agent should narrow your reading burden, not pretend to remove it.
This is where Paperclip is materially different from a generic web-search agent. Its grep can look for an exact pattern across the full corpus, not only titles and abstracts. Its file model can surface the methods section, a figure, or a supplement rather than only the conclusion. Its map and reduce operations are designed for repeated, structured questions across a manageable source set.
For a clinician-scientist, that can be useful in three recurring places:
- Landscape work: rapidly map papers, preprints, regulatory records, and active or completed trials before deciding where a review or grant needs deeper work.
- Evidence extraction: build a first-pass comparison of cohorts, endpoints, interventions, biomarkers, or failure modes across a small set of papers.
- Agent-assisted research workflows: give a coding agent a source-aware tool instead of asking it to hunt for PDFs and then infer what is inside them.
The part I would be careful about
The interface can make extraction feel more authoritative than it is. A structured table is still only as good as the question, the source set, and the paper's reporting. A system may identify that a trial used progression-free survival as an endpoint without telling you whether the trial was powered appropriately, whether assessment was blinded, or whether the population resembles your own.
Use Paperclip to accelerate reading, not replace it
Let the agent retrieve, organize, and compare. Keep the final interpretive work anchored to primary sources, especially when the question turns on study design, endpoint choice, population fit, safety, or a clinically consequential conclusion.
There is also a reproducibility advantage. When you use a documented query, a defined source set, and visible source paths, you can show a collaborator how you reached an answer. That is much better than “an AI said this after searching the web.” It does not make the workflow automatically rigorous, but it gives you the pieces needed to make it rigorous.
Where it fits in a research stack
I would not replace NotebookLM or a reference manager with Paperclip. They solve different problems. A reference manager helps you preserve and organize the papers you already care about. NotebookLM is excellent once you have a curated packet you want to interrogate repeatedly. Paperclip is strongest earlier in the workflow, when you need to discover, inspect, and extract across a live corpus or give an agent grounded access to it.
The practical pattern is: use Paperclip to find and triage, save the core sources to your library, then use a tool like NotebookLM or Paperpile Ask AI for prolonged work with a deliberate collection. If you are building an agentic workflow, keep the same boundary: retrieval and structured extraction can be automated; interpretation, synthesis, and decisions still need a researcher who understands the field.
How we use it
In our own tech stack, Paperclip is the source-grounding layer for research that needs to move faster than a browser-tab workflow. We use it to retrieve and inspect the original papers, trials, and regulatory records behind a question, then carry the relevant sources into the next tool or the final piece of writing.
That includes newsletter work. When we prepare Beam Notes and other research updates, Paperclip helps us verify the primary source behind a claim, check the details that disappear in secondary coverage, and keep the paper or trial record close while we decide what is worth sharing. It does not write the editorial judgment for us. It gives us a cleaner path from a signal to the material that supports it.
That is what Paperclip embodies. It is not an AI replacement for literature review. It is a more capable substrate for doing literature review with agents, while keeping the paper, the figure, the method, and the trial record close enough to inspect.