The document workbench

AI that works across all your documents, not one at a time.

Abstract opens your whole project - research, drafts, reports, policies - and an agent works across all of it. Every change arrives as a proposal you approve, one diff at a time.

Early access · free while in beta · no credit card
The Abstract workbench: a project of files in the left rail, a document open in the editor, and an agent working across the project - applying updates and holding one change back for approval.

Works with the files and sources your work already lives in

MarkdownPDFtranscriptsCSVGoogle SheetsSalesforceREST API
The problem

Your documents live where the AI cannot really work on them.

Chat tools treat a document as an attachment: regenerated whole, edited on faith, one at a time. A page and a blank cursor give you the other half, with the AI bolted onto architecture from another era. Meanwhile the actual work is a folder - the research, the transcripts, the drafts, the report, the summary - and it all drifts apart. A number copied out of one file goes quietly wrong the moment the file behind it moves on.

What only Abstract has

Built on three things no editor does.

Project-scope agency

The agent works across your whole project.

Open a folder and the agent can read and edit every file in it. One instruction fans out across two dozen policies, or keeps a strategy doc, its summary and its outline saying the same thing. No other document tool works past a single file.

More on the workbench
A project-wide agent run: one instruction reads every document in the project in parallel and proposes changes across the ones that need them.
Every change is a proposal

Nothing lands that you did not approve.

Every edit arrives as a diff you can accept, tweak or reject, one at a time or all at once. Change a single clause of a contract and trust that nothing else moved. Where a document runs on real numbers, the rule gets sharper still: figures auto-apply, meaning waits for you.

More on diffs and trust
Cross-document review: each proposed change is shown in context with its source and a confidence label, accepted or rejected one by one.
Living documents

Documents that keep themselves current.

Put an agent on a schedule, a file change or a command, and a document updates itself as its sources move. The weekly report assembles overnight and waits for your sign-off in the morning. Better than a dashboard, because it carries the story, not just the numbers.

More on agents and living documents
The project home: overnight agents have refreshed documents across projects, with the changes that need approval gathered in one place.
One instruction, every document

It scales past a single document.

Point an agent at a whole project - a two-dozen-document policy set, a research corpus, a client folder. Sub-agents read every file in parallel and propose changes only where the instruction actually bites.

A project-wide agent run: 24 sub-agents read every ISMS document in parallel, proposing 38 changes in 12 documents.

One instruction fans out across 24 documents in parallel.

Review everything in one pass.

The proposals gather into a single review. Work through them document by document; each change carries its source line and an honest confidence label.

Cross-document review: changes are accepted one by one, each labelled with its source line and confidence.

Every change carries its source and confidence. Accept, tweak or reject.

One glance answers: what needs me?

Agents work while you are elsewhere. The project home gathers everything waiting on your judgement, across every project, into one calm queue.

The project home: 13 changes need approval across 5 projects; 4 agents ran overnight.

13 changes need approval across 5 projects; 4 agents ran overnight.

How it works

Three steps, then it maintains itself.

01

Open your project.

A folder of real files on your machine - research, drafts, transcripts, data. No migration, no import.

02

Put the agent to work.

Ask for edits across one document or all of them, and @-mention exactly the context it should use.

03

Approve what changes.

Every edit is a diff. Routine updates apply silently where you allow it; anything that changes meaning waits for you.

The flagship use case

When the document runs on real numbers.

This is where Abstract started, and it still goes deepest here. Bind every figure to its source, and the agent drafts around numbers it cannot make up. Figures auto-apply. Meaning waits for you.

The source drawer: clicking a bound figure opens the exact spreadsheet row behind it.
Every number knows where it came from. Click a bound figure and the source drawer opens on the exact row behind it, with a full trail of what changed and when.
Trust controls: figures auto-apply, meaning changes wait for review, and confidence is labelled honestly rather than faked as a percentage.
A number ticking up updates silently and reversibly. A rewrite or a new claim pauses as a proposal, so nothing lands that you did not see.
Why not just…

You've tried the workarounds.

Why not just… Google Docs or Word?

One document at a time and a blank page. AI is bolted onto architecture from another era, with no project around it and no way to work across your files. Abstract sits alongside them, and exports back out.

Why not just… ChatGPT or Claude?

The chat is first-class and your document is the attachment - regenerated whole, edited on faith, with no project. Abstract makes the document first-class and every edit a diff you can see.

Why not just… Notion?

A workspace that holds pages, not a workbench that works across them. Your documents live in its database; in Abstract they stay portable files on your own machine.

Why not just… a BI dashboard?

Great for charts, silent on the story. Stakeholders read the narrative, not the dashboard. A living document carries the numbers and what they mean.

A note from the founder

My work is folders of documents. Strategy, research, product - the real thinking lives across a dozen files at once, not in any single one of them.

The only tool that could actually work across all of them was a code editor built for engineers. Everything made for the rest of us stops at one document and a blank cursor. So I kept doing by hand what the AI was clearly capable of doing across the whole project.

Abstract is that capability rebuilt for people who will never open a code editor. Your project is a folder of real files, an agent works across all of them, and every change is yours to approve.

We're letting people in gradually so we can work closely with early users. If your week lives across more than one document, I'd love you to try it.

TomFounder, Abstract
Joining founders, strategists, consultants and operators already on the list
FAQ

The questions you're about to ask.

What kinds of documents does it work with?

The serious documents you build from source material - strategies, research, reports, policies, briefs, client work. They are plain Markdown underneath, and you can export to DOCX, Google Docs or PDF, or publish to the web.

How does the agent know what to work on?

You point it. Project instructions, a knowledge base of plain files, and @-mentions tell it exactly what to read, and a context panel shows precisely what it is seeing before it does anything.

What can the agent change, and what can it not?

It proposes; you approve. Every edit arrives as a diff you accept, tweak or reject. You can let some documents apply routine updates on their own, and anything that changes meaning always waits for you.

Do I need to be technical?

No. It looks and feels like a document tool, your files stay ordinary files, and you never have to open a terminal or edit a config to use it.

Where do the numbers come from?

Files like CSV and Markdown, plus connected sources such as CRMs and APIs. Your systems stay the source of truth, and in a report every figure is bound to the exact row, field or endpoint behind it.

Is my work locked in?

No. It is plain, portable Markdown you own and can open in any editor. Nothing is trapped inside Abstract.

Can my team review and version documents?

Yes. Every change is tracked with its source and who approved it, so you always have a full history of what moved and why.

What will it cost?

Free while in beta. Paid plans will be per-seat with usage on top, and early users get preferential pricing.

Your whole project, one workbench.

If your week lives across more than one document, we'd love you to try it. Join the early-access list and we'll let you in as soon as your spot opens.

Early access · free while in beta · no credit card