Free tool
Trellis
A small Mac app that builds the accessibility structure into a PDF that was exported without any, and asks you about the parts that are genuinely a judgement call rather than pretending it knows.
Formerly PDF Tagger. Same app, better name: it gives a flat PDF the structure a screen reader can follow, the way a trellis gives a plant something to climb.
A PDF is the easiest thing to share and the easiest thing to break
Exporting to PDF feels like the safe option. It looks the same everywhere, nobody can nudge your layout, and you can email it to anyone. That is exactly why so much gets sent that way, and exactly why so much of it is unusable for anyone relying on a screen reader.
A PDF stores where each piece of text sits on the page. It does not, by default, store what any of it is. You look at a line set in 28pt bold at the top of a slide and you know it is the title. A screen reader has no size to look at. It needs the file to say, in a structure tree, "this is a heading, and this is the paragraph under it, and these three things are a list". Without that tree there are no headings to jump between, no lists, no table rows, no descriptions for the images, and no reliable order to read it in.
Export a deck from Google Slides and you get exactly that: a file with no structure tree at all. Not a thin one, not a wrong one. None. The same goes for plenty of other exporters. Google Docs has a worse habit, which is exporting tags that exist but are nonsense, so checking tools report the document as tagged and everyone moves on.
Conformance is not the same as usable
While building this I tagged a real deck in which roughly three quarters of the text came out wrongly marked as headings. It passed all 106 PDF/UA rules. Every one. An automated checker can tell you a heading exists; it has no way to tell you the thing you called a heading is a sentence from the middle of a paragraph.
Why it does not try to do the whole job on its own
It would be easy to make this fully automatic, and the result would be worse. A program can measure a font size, a position, the gap between two lines. There are limits to what that tells you:
- Is that bold line a heading, or the first few words of a sentence? Plenty of decks set the opening word of a paragraph larger and bolder. Call it a heading and you have put half a sentence into the list a screen reader user navigates by.
- Are those two columns a table, or two cards sitting side by side? Geometrically they are identical. One should be read across the rows, the other one card at a time, and reading a pair of cards as a table shreds both of them.
- Does that image mean anything? Only the person who put it there knows whether it is a chart carrying the argument or a swoosh behind the title. Getting this wrong in either direction is bad: a described decoration is noise, an undescribed chart is a hole in the document.
So the app measures everything it reasonably can, proposes a role for each item with a confidence score and a plain sentence explaining why, and puts the genuinely uncertain ones in a queue for you. You confirm, correct, merge, split, or drag things into a different reading order. The aim is to take the tedious ninety percent off your hands and be honest about the rest, not to produce a confident-looking file nobody checked.
What it actually builds
- Headings at sensible levels, with lists, quotes and captions kept out of them.
- Real tables, including borderless ones, with header cells scoped to their row or column and the reading order running across the rows.
- Figures with alt text you write, or mark as decorative so they are skipped.
- Links and form fields, labelled from the text next to them.
- A reading order you can see, click through against the page, and drag into place.
- Document language, title, and the flags a reader needs to show the title rather than the filename.
- A named list of anything it could not attach a tag to, before you export and again in the result, rather than a file that quietly lost a heading.
It checks its own work with veraPDF against all 106 PDF/UA-1 rules. That is bundled inside the app with a trimmed Java runtime, which is most of the download size, and the reason you get a real verdict instead of a promise. Everything happens on your machine: it is a local app with a local interface, and the only thing that ever leaves your Mac is whatever you choose to send.
The one optional exception is a switch in the top right. If you use Claude Code, turn it on and Trellis will draft a description for each image and give the reading order a second look. It works out whether an image carries meaning at all or is just decoration, and tells you why it decided that, which is usually the part people find tedious. Everything comes back as a draft that still counts as outstanding until you have read it, and the app is perfectly usable with the switch off.
It never touches your pixels. Open the tagged file next to the original and they are identical; all that changed is the structure underneath.
Installing it
- Download the zip and unzip it.
- Drag Trellis into your Applications folder.
- Double-click it. The first time, macOS will refuse and say it cannot check it for malicious software. Open System Settings ▸ Privacy & Security, scroll to the bottom, and click Open Anyway next to Trellis. You only do this once; it opens normally from then on.
Why macOS complains
The app is not signed with a paid Apple developer certificate, so Gatekeeper treats it as coming from an unidentified developer. Open Anyway in Privacy & Security is the way past that on current macOS; the old right-click trick no longer works. If you would rather not, the source is on GitHub and it runs from a Python environment in a couple of commands.
Using it
- Open a PDF. The app reads it and proposes a structure.
- Walk the pages. Each item shows what it was called, how confident that was, and why. Anything uncertain is collected into a review list so you are not hunting for it.
- Fix what is wrong. Change a role, merge two fragments, split one that swallowed two columns, or drag an item to a different place in the reading order.
- Describe the images, or mark them decorative. With Claude Code switched on, the app can draft these for you first.
- Export. You get a conformance verdict on the way out, and the name of anything that could not be tagged, because the verdict alone cannot see that.
There are two short guides in the app's downloads, one on installing and one on how it works. Both are, as you would hope, tagged.
Getting a hand from Claude Code
Two jobs in this app are beyond a script: writing a description for each image, and judging whether a page reads in a sensible order. If you already use Claude Code, Trellis can hand both to it. There is no key to paste and no separate account: it runs on the Claude Code sign-in you already have, so it counts towards that account's usage like anything else you do there.
Setting it up
-
Install Claude Code, if you have not already. In Terminal:
curl -fsSL https://claude.ai/install.sh | bash, orbrew install --cask claude-codeif you use Homebrew. It needs a paid Claude plan (Pro, Max, Team or Enterprise) or an Anthropic Console account; the free plan does not include it. - Quit and reopen Trellis. It looks for Claude Code when it starts. Once it finds it signed in, a Claude off switch appears in the top right, next to Help tips.
- If it says Claude Code is not signed in, press Sign in to Claude on the first step. A Terminal window opens Claude Code's own sign-in, your browser asks you to log in, and the switch appears by itself once you are done. Claude Code keeps its own sign-in, separate from Claude in the desktop app or a browser, so this can happen even if you use Claude every day.
Using it on a document
- Flip the switch to Claude on. Nothing is sent yet. It stays on until you switch it off or quit, and it is never remembered between launches, so every session starts with it off.
- Draft the image descriptions. On the Images step, press Draft the missing descriptions (the button gives the count). A bar shows how far it has got; on a long deck it can take a few minutes. For each one, Claude decides first whether it is a logo, decoration, a chart, a screenshot or a photo. Decoration is marked to be skipped; everything else gets a one-sentence draft, with Claude's reason shown on the item.
- Read every draft against its picture. Drafts arrive marked unread and low-confidence, and they count as outstanding until you have looked. That is deliberate: a description that reads well and describes the wrong thing passes every automated check and helps nobody.
- Check a page's reading order. On the Structure step, press Check this page's reading order under the page. Claude lists anything that looks out of sequence and why; click an item to jump to it. It is a suggestion, not a change: you drag things into place yourself.
What leaves your Mac
- Nothing while the switch is off, and nothing when it is on until you press one of the two buttons above.
- For a description: a crop of that one image, plus up to 600 characters of the text on the same page so Claude has some context. Not a picture of the page, and nothing from any other page.
- For the reading-order check: the text of that one page (each item cut to 90 characters), what each item is tagged as, and where it sits. No picture at all.
- It goes to Anthropic through your own Claude Code account, on whatever terms that account has. For a document you would not paste into a chat window, leave the switch off and write the descriptions yourself.
If it does not work
-
“Claude Code is here, but not signed in”. Trellis found
it, but the terminal command is signed out. Press Sign in to Claude
and finish signing in; the switch appears without a restart. If Terminal does not
open, run
claude auth loginin Terminal yourself, then press Check again. -
No switch and no message. Trellis did not find Claude Code. It
checks the places Claude Code normally installs itself. Run
claude --versionin Terminal: if that fails too, Claude Code is not installed. If it works, quit and reopen Trellis, since it looks when it starts. - Your sign-in lapses part way through. The switch goes off and a Sign in to Claude button takes the place of the one you pressed. Sign in, switch Claude back on and press it again. Any descriptions drafted before it stopped are kept.
- “Claude Code did not answer in time”. Each image gets two minutes. Usually a slow connection; try again.
-
Using an API key instead. The downloadable app only works through
Claude Code. If you run Trellis from source, install the
anthropicPython package into its environment and setANTHROPIC_API_KEYbefore starting it. If Claude Code is also installed and signed in, Trellis uses Claude Code.
The honest caveats
- This is a personal project and a work in progress. It is built in my own time, it has rough edges, and it is not a product. I use it and it does the job, which is a different claim from it being finished.
- Apple silicon Macs only, and not signed or notarised. See above.
- Review the output. The app is careful about flagging what it is unsure of, but it still proposes and you still decide. Do not ship a document you have not looked at.
- Tagging does not fix a bad document. If the text is 3:1 against a photo, or the content only makes sense if you can see the layout, structure will not save it. Tagging makes a good document reachable; it cannot make a poor one good.
- The real fix is upstream. If a document can live as a web page, it should. HTML gives you structure for free, reflows on a phone, and does not need a tool like this at all. Trellis exists because "just send it as HTML" is not always an option, not because PDF is a good idea.
If you try it and something breaks, or it calls something a heading that plainly is not, I would genuinely like to know. Issues on the GitHub repo are the easiest way, or there is a contact link on the front page.
