No one sits between you and your data.
Mail, calendar, files, documents, spreadsheets, presentations, photos, notes and AI chat — running on hardware you own, or a cloud account you pay for directly. No subscription that locks you out of your own files when it lapses. No service that can be shut down. No terms that change on a Tuesday.
Free for personal use. No account. Nothing to cancel.
Shipping · coming
Every dollar you spend feeding documents to an AI becomes 8 cents.
A language model never sees your document. It sees your document wrapped in every style definition, revision marker and XML envelope the file format drags along — and you pay for all of it, on every single request. Convert to GFM+ first and the wrapper is gone; the content, the tables, the headings and the footnotes are not.
.docx
Same content. Same structure. Converts back to .docx when you're done —
59/59 documents round-tripped with zero semantic differences.
.md, a real .xlsx, a real
.pptx — open side by side in one editor, all GFM+ underneath. This is a
screenshot of the actual app.
See the full measurement — including where it goes against us →
Stop renting your office software.
Ten apps that share a format, a storage layer and a component library — which is why they feel like one product instead of a bundle. Everything below is a screenshot of the real thing, and the editor is running live further up this page.
.xlsx, recalculated and editable in
a live grid, in your browser, from your disk.
OUI EditorLive here
Long-form writing. WYSIWYG or raw Markdown, your choice, toggled instantly.
replaces
OUI OfficeBeta
Documents, spreadsheets, presentations. Opens and saves real Office files.
replaces
OUI FilesBeta
One file manager for every backend you've got. Speaks WebDAV.
replaces
OUI MailBeta
JMAP mail. Point it at Fastmail, at your own server, or at any IMAP box through the proxy.
replaces
OUI CodeBeta
An AI coding workspace, not a scrollback buffer — driving the subscription you already pay for.
replaces
OUI NotesBeta
Linked, tagged, searchable notes. Plain Markdown files on disk.
replaces
OUI ChatAlpha
Every model you use, in one place, with your work as context.
replaces
OUI CalendarComing
Your calendar, on your server, sharing an account with Mail.
replaces
OUI ContactsComing
Your address book, in a format you can read without our software.
replaces
OUI PhotosComing
Import, organise, view. Reads HEIC and EXIF, finds duplicates, thumbnails server-side.
replacesStatus is on every card and every app page carries a dated list of what it does not do yet. We would rather you found out here than three weeks in.
Every other page you visited today is measuring you.
This one can't — there is nothing here to do it with. No analytics SDK, no tag manager, no font CDN, no cookie banner, because there are no cookies to consent to.
The bottom row is deliberately not zero. This page loaded some HTML, some CSS and four small scripts, and pretending otherwise would be the exact dishonesty the counter exists to refute. What is zero is everything above it: nothing here contacted a third party, set a cookie, or sent a byte about you anywhere.
The numbers come from your browser's own PerformanceObserver, not from us —
and they include any request the live editor makes when you load it. Open your network
tab if you would rather not take our word for it. That is the entire point of putting
this here.
Measured, not asserted.
Three axes — token cost, structure retention, round-trip fidelity — over two corpora, one of them built by people trying to break a parser. Including the results that go against us, because a benchmark that only reports its wins is an advertisement.
fewer tokens than the .docx it came from
of document structure survives plain-text extraction
documents round-tripped with zero semantic differences
A pipeline has three options, and only three.
Put the OOXML in the context window — structure-complete, and enormous. Extract the text — cheap, one-way, and it discards every table, heading, footnote and merge, which is usually what the question was about. Or convert. Here is what each one costs:
| Corpus | As OOXML | As GFM+ | Ratio |
|---|---|---|---|
| Documents58 files | 1,092,496 | 86,228 | 12.7× |
| Presentations28 decks | 3,423,895 | 632,629 | 5.4× |
| Spreadsheets15,870 workbooks | 4,819,304,989 | 1,069,300,640 | 4.5× |
Range across a representative corpus and an adversarial one: 3.7× – 16.6× — across a representative corpus and an adversarial one built by people trying to break a different implementation.
Cheap is not the same as free.
Text extraction is cheaper than GFM+. If tokens were the whole argument, extraction wins. So it is worth being exact about what it keeps:
| What the document had | Kept by text extraction | Kept by GFM+ |
|---|---|---|
| Tables161 in source | 0.0% | 100.0% |
| Table rows491 in source | 0.0% | 100.0% |
| Table cells1,437 in source | 0.0% | 99.2% |
| Headings78 in source | 0.0% | 92.3% |
| Footnotes6 in source | 0.0% | 100.0% |
| List items629 in source | 0.0% | 97.8% |
Not "little." None. A plain-text extraction preserves 0.0% of every structural construct, whichever extractor you use, because none of them carries a grid, a heading level or a footnote binding. Its token advantage is not a saving — it is the price of the answer.
And it converts back. That is the part nothing else does.
Anyone can produce a smaller representation of a document. One that converts back — into the file the world actually wants to receive, with styles as styles, comments as comments and tracked changes as revisions — is infrastructure rather than a preprocessing trick.
| Axis | Representative corpus | Adversarial corpus |
|---|---|---|
| FidelitySomeone else's .docx → GFM+ → .docx, compared semantically | 59 / 59 | 28 / 28 readable |
| StabilityOur GFM+ → .docx → GFM+, compared byte for byte | 59 / 59 | 38 / 38 |
Where these numbers come from, and where they are weakest
Every figure on this page is reproducible from the benchmark harness. We publish the corpora, the licences, the tokenizer and the failures — including the four places the current report corrects its own predecessor, one of which narrowed our own advantage.
- docx
- 58 files — US government forms (17 USC 105), MIT/Apache/BSD test suites from other projects, and synthetic single-construct fixtures. Each recorded with source URL, licence, retrieval date and sha256.
- xlsx
- 15,870 of 15,871 public Enron workbooks. The one exclusion is named in the report: 353 million characters of XML that the tokenizer cannot count.
- pptx
- 28 of our own decks.
- tokenizer
- gpt-tokenizer, cl100k/o200k family.
- report
- gfmplus-benchmark-report-v0.0.11-LnxAby, measured 2026-08-11
Two things this does not establish, stated here rather than discovered later: models do not demonstrably reason better over GFM+ — on questions where both formats contain the answer the difference is a coin flip. The advantage is that GFM+ still has the fact when extraction has thrown it away. And on decks GFM+ currently answers worse on "which slide contains X", because the deck body does not number its slides. That is a real defect with a known cause and a known fix.
The full measurement, with the methodology and the corrections →
Bring your own key. Or no key at all.
OUI Chat runs local models in the browser through WebLLM, and connects to any remote provider with a key you paste in and we never see. Your documents, mail and files are available as context without leaving your infrastructure — the request goes from your browser straight to the provider, because there is no "us" in the path.
Alphabetical. No hierarchy, no default, no revenue share.
Your hardware, your cloud account, or both.
Your hardware — Undercast
A single binary you install on a machine you own. No runtime to install first, no platform account, no other services to stand up alongside it. It mounts as an ordinary drive in Finder, Explorer, Nautilus and Dolphin, because underneath it is WebDAV and JMAP.
Your cloud — Cloudflare
The same protocols against R2 and Workers in an account you hold and bill directly. Your data sits with a provider you chose and can leave. The apps do not know or care which backend they are talking to.
Software you run, on infrastructure you control.
That's the whole idea. Everything else is implementation.