Alternative to Connected Papers
Connected Papers vs. Agent Bayes: Which Fits Your Workflow?
Connected Papers finds what to read. Agent Bayes reads it and builds a citation-backed map of what it says. Two stages of the same workflow, not two options.
Choose Connected Papers if
- You are new to a subfield and need to see its shape in ten seconds.
- You have one seed paper and want to find its intellectual neighbourhood.
- You want to spot foundational prior work and recent derivative work quickly.
- You want something that costs almost nothing.
Choose Agent Bayes if
- You already have the papers and need to know what they say and where they disagree.
- Claims need author, year, and printed page range for a bibliography.
- You want a workspace that grows over months, not a graph regenerated per query.
- You need to search inside your papers, including footnotes, tables, and appendices.
- Your library includes papers in languages other than English.
Connected Papers and Agent Bayes at a glance
| Axis | Connected Papers | Agent Bayes |
|---|---|---|
| Question it answers | Which papers are related to this one. | What do the papers I chose actually say, and can I defend it. |
| What it reads | Metadata, references, citations, and abstracts from the Semantic Scholar index. It does not read full text. | Every page of every PDF in your knowledge bases, with layout-preserving OCR. |
| How connections are made | Co-citation and bibliographic coupling, so two connected papers need not cite each other. | Semantic retrieval over indexed passages, plus a terminology graph built from your own corpus. |
| Output | A force-directed graph generated per seed paper. | A mindmap of cited claims you and the agent develop, with version history and snapshots. |
| Citations | Nodes are papers. There are no claims to cite. | Author, year, printed page range, and the exact source chunk, opened in a built-in PDF reader. |
| Does the work accumulate | A graph is regenerated from its seed. Nothing you wrote lives in it. | Projects hold several mindmaps with conversation history, undo and redo, and named snapshots. |
| Disagreement | Not represented. A graph edge does not say whether two papers agree. | Preserved as sibling nodes, each with its own evidence chain. |
| Pricing | Free tier with 5 graphs per month. Academic around 6 USD per month and Business around 20 USD, both billed annually, as of 3 August 2026. | Metered credits charged in 0.25 increments against an inspectable ledger. |
What Connected Papers gets right
Enter one paper, get a map of its intellectual neighbourhood in seconds. For orienting yourself in an unfamiliar subfield there is still nothing faster, and the free tier of 5 graphs a month covers occasional use at no cost.
The mechanism is the interesting part and it is why the tool works. Edges are not citations. Similarity is computed from co-citation and bibliographic coupling, drawn from the Semantic Scholar index, so two papers can be strongly connected without either one citing the other. That is exactly what you want when you are new to an area, because it surfaces conceptual neighbours that share no keywords with your search terms. Prior Works shows the earlier papers most cited by the graph, and Derivative Works shows the later papers citing many of them, which gives you the foundations and the frontier in two clicks.
Paid plans remove the monthly cap for around 6 USD a month on the academic tier as of 3 August 2026, which is close to free.
Where it stops
Connected Papers never opens a paper. It works from metadata, references, citations, and abstracts. That is not a shortcoming of the implementation, it is the design, and it means the tool cannot tell you anything about what a paper argues, what evidence it presents, or whether two papers in the same cluster reach opposite conclusions.
A graph edge does not encode agreement. Two papers sitting next to each other on the map may be in direct opposition, and the layout will not show it.
Nothing you learn goes into the graph. You cannot write a claim into it, attach evidence to it, or come back in a month and see what you concluded. A graph is regenerated from its seed, so it is a view of the index rather than a record of your thinking.
And you still end up with the real problem. You now have 80 papers you have not read and no structure to put them in.
What happens next, in Agent Bayes
Agent Bayes starts at the point Connected Papers stops.
It reads. Every PDF runs through OCR that preserves multicolumn text, tables, figure captions, footnotes, and headers, then semantic chunking, then distillation into structured bullet points that carry enough context to stand alone. Those bullet points are always written in English regardless of the source language, so a French or Japanese paper is searchable next to an English one.
The map holds claims, not papers. The agent writes cited claims into a mindmap and you shape it: rewrite any node, restructure branches, attach or remove citations, embed images, and pin nodes as context for the next instruction.
Every claim has a page. Author, year, printed page range, and the exact chunk of source text, opened in a built-in PDF reader at that page. Complex syntheses carry several citations on one claim.
Disagreement is preserved. When your sources conflict, competing positions become sibling nodes with separate evidence chains rather than one smoothed statement, and you can ask for a cited prose discussion of where the evidence diverges.
Work accumulates. Projects hold several mindmaps, each with conversation history, a rolling version stack for undo and redo, named snapshots, and provenance recorded on every change.
There is a concept map too, of a different kind. The terminology graph extracts domain-specific terms from your own documents and connects them by how often they appear together in the same passage. Search a term, explore its neighbourhood, and click any connection to read the passages behind it. The research agent consults it to resolve what an evolving or ambiguous term means in your specific field. Building it is free.
Alternative to Connected Papers, or the stage after it?
Map the field in Connected Papers. Save what matters into Zotero. Index it in Agent Bayes with the Zotero plugin, which works on your items in place with no re-uploading and shows the estimated cost before you commit. Then build the argument, with every claim traceable to a page.
Agent Bayes does not search the open literature and will not find you papers you do not have. That is what Connected Papers is for.
Questions researchers ask about Connected Papers
Is Agent Bayes a Connected Papers alternative?
No, and you should be suspicious of any page that claims otherwise. Connected Papers does discovery over a public citation index. Agent Bayes does synthesis over the library you built. If you replaced Connected Papers with Agent Bayes you would lose the ability to find papers you do not have.
Does Agent Bayes build a citation graph like this?
It builds something different. The terminology graph extracts domain-specific terms from your own documents and connects them by how often they are discussed together in the same passage. You can search a term, view its neighbourhood, and click any connection to read the exact passages behind it. The research agent also consults it to work out what a polysemous term means in your field. It maps concepts in your corpus, not papers in an index.
How do I get from a Connected Papers graph to Agent Bayes?
Save the papers you want into Zotero, which Connected Papers exports to. Then use the Agent Bayes Zotero plugin to index them in place, with no re-uploading and no duplicated metadata. The plugin shows the estimated indexing cost before you commit.
Why does Connected Papers connect papers that never cite each other?
Because similarity is computed from co-citation and bibliographic coupling rather than direct citation. Two papers that share many references, or that are frequently cited together, are treated as related. That is why it surfaces conceptual neighbours that keyword search misses, and it is a genuine strength.
Can I use both?
Yes, and this is the recommended pairing. Map the field in Connected Papers, collect what matters in Zotero, index it in Agent Bayes, and build the argument there.
Bring your own papers and see what they actually say
Index your library, ask the agent to map it, and open any claim at the page it came from. Zotero users can send papers across without re-uploading anything.