abstract searches the literature, reads your papers, and drafts with you, and every cited sentence is checked against the passage it cites. Free, open source, and it runs on your own computer.
macOS and Linux. Nothing else to install.
Most AI writing tools ask the model to be careful. abstract checks the model's work with code. Point at a sentence to see what backs it up.
The Transformer replaced recurrence and convolution with attention alone[1]. Residual networks are easier to optimize and always more accurate at greater depth[2]. Residual connections were first introduced for machine translation[2].
verbatim from the source
We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.
Vaswani et al. (2017). Attention Is All You Need
Every part of the claim follows from the passage, so the sentence is marked supported.
verbatim from the source
these residual networks are easier to optimize, and can gain accuracy from considerably increased depth.
He et al. (2015). Deep Residual Learning for Image Recognition
“Always more accurate” is stronger than the source’s “can gain accuracy”, so the sentence is only partly supported.
No supporting passage found.
He et al. (2015). Deep Residual Learning for Image Recognition
No passage in the cited source supports this claim. Export stops until you fix it or explicitly accept it.
Titles, authors, venues, and DOIs are copied from Crossref and OpenAlex records.
A sentence is only marked supported if the checker quotes the source word for word, and code confirms the quote is really there.
Export re-checks every DOI for resolution and retraction, and stops on unsupported claims unless you accept them.
One agent plans the work, shows you its plan, and keeps going until it's done. You can steer it at any point.
In plain words, in a chat. Point it at a folder of your own papers, or start from nothing.
It searches OpenAlex, Crossref, and arXiv, follows citations, and logs every include and exclude decision.
It fetches open-access PDFs, reads them page by page, and keeps notes. Only what it opened can be cited.
It drafts in Markdown and checks every cited sentence and table cell against the passage it cites.
Markdown, LaTeX, and BibTeX, plus an audit file that maps every claim to its evidence.
try: “Review recent work on retrieval-augmented generation and write a related-work section with verified citations.”
Searches OpenAlex, Crossref, and arXiv at once, deduplicated, and follows citation chains backward and forward.
Every search and include/exclude decision is recorded with its reason, and exports as a PRISMA-style flow.
Only sources the agent actually opened can be cited. Downloading a paper is not reading it.
Image-only PDFs are transcribed by your model; figures, tables, and equations become searchable.
Comparison tables and extraction matrices are checked cell by cell, exactly like sentences.
Citations as [1] or (Hooker et al., 2020), with bibliographies built only from registry records.
For big reading lists the agent hands work to up to four sub-agents, under the same integrity rules.
Venue review rubrics (NeurIPS, ACL, CVPR, IEEE, Nature…), methods checklists, and writing guides. You can add your own.
A visible plan, messages the agent reads at its next step, and a stop button that actually stops.
What you ask it to remember is saved right away; preferences it only infers wait for your approval.
Executes Python scripts on your data in the workspace, without access to your API keys.
Checks every DOI in a manuscript for resolution and retraction before you submit.
Pick one model in the chat box and abstract uses it for everything. The list shows only recent models from what you've connected.
Click Continue with ChatGPT and abstract runs on your Plus or Pro plan. Set a usage cap any time in ChatGPT settings.
Paste a key in Settings or use an environment variable.
Run open models on your own machine, so document content never goes to a cloud provider.
Semantic library search uses embeddings from OpenAI, Google, or Ollama when one is connected; otherwise abstract searches by keyword.
~/.abstract. No account, no server of ours.Install in one line. No account, no subscription.
macOS and Linux. Nothing else to install.
Every release is built by GitHub Actions from the public source, and the installer checks its checksum before installing. How to verify