A library for you
and your agents.
Your papers, PDFs, proofs and repositories, indexed on your own machine and readable by both of you. Your agent asks it for the page rather than recalling one, and every answer comes back with the source it came from. Nothing is uploaded.
1Install
$ curl --proto '=https' --tlsv1.2 -sSf https://adduce.dev/install.sh | sh2Tell Claude Code about it
$ claude plugin marketplace add https://api.adduce.dev/releases/marketplace.json$ claude plugin install adduce@adduce
Registers the tools and the backed-research skill together. Cursor, Zed and anything else speaking MCP run adduce-daemon --mcp instead.
Fourteen days, no card.Create an accountOther install routes
A model with no library invents one.
Asked for a source it has never read, a model produces one shaped like the paper you wanted. The citation is well formed, the authors are plausible, and the work does not exist. A study of every paper at ACL, NAACL and EMNLP across 2024 and 2025 found nearly 300 of them citing work that does not exist, most from 2025. Sakai, Kamigaito and Watanabe, ACL 2026.
Your agent has never read your sources.
The papers are on your disk and the model is working from what it absorbed in training. It writes around the gap, fluently, and the citation it produces resolves to a paper that says something else.
A context window contains what you paste into it.
What you paste is what it knows, so the answer depends on which four papers you happened to attach. The other two thousand may as well not exist.
The source moved after you saved it.
It was retracted, corrected, or replaced by a later version, and nothing that rests on it knows.
Fourteen hold. Three could not be checked. One is flagged.
You get a count with its gaps named, and the three or four sentences worth your afternoon. Knowing where not to look is most of what you are buying.
The global electric-vehicle fleet passed 40 million cars in 2023, and adoption is still accelerating. China alone accounted for roughly 60% of new EV sales that year. Battery pack prices have fallen by more than 80% since 2013, which is the single biggest driver of the shift. By 2030, EVs are expected to make up over 90% of all new car sales.
Needs a source. No published forecast supports 90% by 2030. The closest figure Adduce found projects a far lower share. Revise the number or add a citation.
Could not be checked. The cited annual report is a scan with no text layer, so the passage was never read. This claim is neither backed nor flagged.
market-brief.md · 18 claims
| 14 | backed to the page | |
| 3 | could not be checked | |
| 1 | flagged, source superseded |
Why three could not be checked
- a cited annual report is a scan with no text layer
- one source is metadata only, held behind a publisher paywall
- a numerical artefact has not rerun since its script changed
An unchecked claim is never reported as a passing one. One of those would cost the other fourteen their meaning.
One thing to set up. Your agent does the rest.
Install it, and point your agent at it
The two commands above. The first puts Adduce on your machine and hooks your shell; the second registers its tools and the backed-research skill with Claude Code. Adduce then runs as a local service that Cursor, Zed and any other tool speaking MCP can call the same way, and the contract shipped beside the binary tells your agent the rest.
Your agent fills the library
Give it a DOI, an arXiv id or a folder of PDFs. It fetches the paper, reads the mathematics out of the page rather than around it, and indexes the text on your own machine. Your repositories go in the same way, so the code is as readable as the papers.
Both of you read from it
Your agent asks the library for the passage instead of recalling one, and gets back the text with the page it sits on. You search the same library, in the same words, and get the same page.
- Every answer arrives with its page
- The library returns the passage and where it sits, so an assertion can be followed to the source rather than taken on trust. What it could not read it reports as unread, and the third state is never rounded into the first.
- It notices when a source changes
- What you saved is rechecked against the publisher. A retraction, a correction, a new version or an altered title comes back as a finding on everything resting on that source, months after you shelved it.
- Papers, proofs and your own code
- A page of a paper, a machine-checked proof in Lean 4, Z3 or SymPy, or a script that reruns to the number printed in the text. Your repositories are indexed alongside them, so a question about the code is answered by the code.
- The library stays on your machine
- Documents, OCR and embeddings sit on your disk and are read there. What leaves the machine is a request to a publisher for the metadata of something you already have.
Every command answers in the same shape.
Your agent never reads this page. It reads one answer per command, in one shape, with a verdict, a list of findings, and a code on each finding that keeps its meaning between versions. A script can branch on the exit code alone, one command reports whether the install is ready to work, and a long run reports its progress as it goes.
$ adduce claim verify --all total claims: 18 clean: 14 flagged: 1 unknown: 3 overall verdict: flagged
Know what you are signing.
Install Adduce, connect your agent, and get the count before the document goes out.