Your machine holds everything you ever wrote. Now ask it what happened.
Your documents know the terms. Your mail knows the conversation. Your photos know when and where. Your git history knows what changed. Your agent sessions know what it already tried. docenta makes them one private evidence layer, and every answer points back to the source. Nothing leaves your machine.
The revised quote of July 29 is the one that stands: $2,400 below the original and the latest price in the collection. An August 1 email from the contractor says the revised number "stands through September". No later quote or retraction appears in the indexed material.
An illustration of the shape of every docenta answer: the files are made up, the behavior is real. The real ones are on the desk pages below and behind the eight live doors, where any visitor can ask.
Search finds files. You want to know what happened.
Nobody wakes up wanting semantic search. People want to know where the proof is, which version stands, what the record says on a date, and what it does not say. Those are memory questions: several artifacts, a chronology, a source that outranks another, and a way to inspect the proof. That is what docenta is built for.
| What you type into search | What you actually want to know |
|---|---|
| files containing "streaming" | When did we decide streaming, why, and what happened next? |
| emails from the contractor | Which quote replaced the old one, and did they confirm it stands? |
| PDF search for "mold" | Which policy text applied on the date of loss, and what in the notes conflicts with the claim? |
| git blame | Why was this line introduced, and which incident, note or session explains it? |
| papers containing "low dose" | Which papers in my own library agree, which disagree, and on which page? |
| "Palo Alto ADU setback" | For this configuration, under the edition that governs this permit, which setback applies, and do the city and the state agree? |
Local evidence, exact receipts, explicit conflicts and gaps, served to the agent you already use. Built for evidence questions ordinary search misses. A docent is the museum guide you ask about a collection. Your corpus gets one.
The one question that sums it up, and the one a retrieval box bolted onto a chatbot cannot answer: what do my own records actually prove about this, and what do they not prove? It needs the sources ranked, dated and compared, the contradiction kept visible, and an honest "not established" at the end. That is the whole product.
Answers come with receipts. Five kinds.
Gaps disclosed, never papered over. When the corpus cannot answer, docenta says so, shows what it did find, and names the document that would settle it.
Ask the sensitive question without first uploading the sensitive file.
The matter file, the claim, the mail archive, the private source tree, the unpublished paper, the household folder: docenta ingests, indexes and answers on your own hardware. Originals are never copied; only searchable derivatives are stored. Your agent gets a small evidence pack with the receipts, not your archive. It gets it over MCP, the connector standard it already speaks, from a server that runs on your own machine: your agent talks to a local process, not to a URL in someone's cloud. The only remote MCP addresses on this site are the public doors, sealed collections we host on purpose; your own collection never has one. Zero uploads, zero telemetry, and the token budget is printed on every answer.
One engine, your kind of collection.
The same docent, pointed at what your desk actually holds. Each page shows one real question and the shape of its answer; eight public collections are live behind doors today, and every door answer below was asked through the same engine.
Contractors. Ask the rule. Get the exact page.
City code, handouts, state law, utility standards and the code edition tied to the permit date, as one dated collection. Which rule applies, where the exception lives, whether the sources agree, and what the corpus cannot establish.
"Which code edition applies to this permit date?" · "Do the city handout and the ordinance agree?"
Lawyers. Ask the matter. Cite the exhibit.
Pleadings, correspondence, scanned exhibits, contracts and notes become one private evidence layer. Reconstruct the chronology, find the version that changed, surface the contradiction, map each assertion to a page, without uploading the matter anywhere. Try the engine today on the federal code and public contracts.
"Where does the other side first acknowledge this?" · "Does the declaration conflict with an earlier email?"
Insurance. Ask the claim, not the folders.
Policy, notes, mail, forms, estimates and photos as one timeline with receipts: what happened, which version applies, where the file conflicts, what proof is missing.
"Which policy was in force on the loss date?"
Home. Your files remember what you forgot.
Twenty years of mail, documents and photos, asked in plain words: the receipt tied to the warranty, the quote that replaced the old quote, the letter you half remember.
"Is the dishwasher still under warranty?"
Mail. Ask what was said, not what the subject line was.
The promise, the latest number, the attachment that went with it and the whole thread in order, even when you remember the meaning and not the sender.
"What was the last delivery date they committed to?"
Developers. Ask why the code is this way.
Code, git history, design notes and every agent session as one cited memory layer. Test it on a project you know: the Git project, every commit since 2005 and the whole mailing list, sealed as a public collection behind its own door.
"Why is it called --force-with-lease, and what was wrong with the other names?"
Researchers. Ask your library where it disagrees.
PDFs, scans, notes, drafts and co-author mail as one evidence layer: the paper you half remember, the results that conflict, the sentence in your draft that has no support.
"Which papers support this, and which contradict it?"
Eight sealed collections you can ask today. Each with its own card.
Public material only, one collection per subject, never mixed. Every door carries a card sealed with the collection and shown to your assistant on every connection: exactly what is inside, what is not, and the date nothing changed after. Register once on any page and the doors are open to you; each page below has a real answer asked through its door.
Judaism. Tanakh, Mishnah, Talmud and commentary
The Hebrew Bible in the Westminster Leningrad Codex with Open Scriptures morphology, and the openly licensed versions of the Sefaria library: Tanakh, Mishnah, Talmud, Midrash and commentary in Hebrew, Aramaic and English, 183,584 documents, one record per chapter or daf, each cited to its version and license.
"Why does the Torah begin with creation and not with the first commandment, and where does Rashi get that from?"
Christianity. Scripture and classical commentary
The Berean Standard and World English Bibles chapter by chapter, Matthew Henry complete and concise, sixty-three volumes of Spurgeon, the SBL Greek New Testament, the Patristic Text Archive and the First Thousand Years of Greek: 51,876 documents, English first, Greek where the source is Greek.
"What does Matthew Henry make of Psalm 23:4, and did Spurgeon preach on that verse?"
Islam. The Koran in six translations and Arabic, and the hadith
Pickthall, Yusuf Ali, Shakir, Sale, Rodwell and Palmer sura by sura, the Arabic text with and without diacritics, and the ten hadith collections in English and Arabic, every hadith under its number with its gradings. No tafsir; the rights in the hadith texts are stated as unresolved on every record.
"How do the six translations render Surah 112:1, and which hadith say what the surah is worth?"
Faith. The three traditions together
The union of the three collections, 237,496 documents, nothing re-weighted: every answer says how its matches divide among Judaism, Christianity and Islam, and a question can be put to one of them alone.
"Where do the three scriptures tell of Joseph sold by his brothers, and do the commentators agree who sold him?"
Sleep. The open research corpus
77,899 openly licensed texts: PubMed Central full texts under CC0, CC BY and CC BY-SA, ClinicalTrials.gov registrations, one AHRQ evidence report. Research literature, not medical advice; no subscription literature and no abstracts of closed articles.
"Which trials tested melatonin for insomnia in older adults, what did they measure, and does the meta-analysis agree?"
Diabetes. The open research corpus
The same shape as sleep, for human diabetes research. Being gathered; the card is written when the collection is sealed.
door in preparation
A spine that ingests, three channels that answer, one artifact that proves it.
channels exact BM25 · paragraph lexical · dense vectors, fused and re-ranked · entities, timelines, themes on top
answer evidence_pack: route plan, cited spans, conflicts, disclosed gaps, enforced token budget
live mail, chat exports, git and agent sessions stream in every hour · your corpus feeds itself
tools search search_meaning evidence read thread related who_is timeline verify_quote status and ten more, over MCP · every page says what was searched, what was dropped and why, and how many documents held every term
The same 48 judged questions over the same corpus, asked by the same scripted agent two ways. With the evidence pack it cited the right document on 44 of 48, in 1 to 2.5 tool calls per task. Limited to search-and-fetch loops, the way an agent works without the pack, it cited it on 39 of 48, in more calls. The difference is what the agent is handed: not a hit to go and read, but the evidence itself, the cited spans, the dates, the conflict and the gap, assembled and ranked, so the first answer is the deep one. What each pack used is printed on it. The lexical, entity and timeline layers are deterministic: a re-run gives the same bytes.
Everything on the machine, by what it is.
One engine, every lane. Originals stay where they are; docenta keeps searchable derivatives and a map back to the path. The list below is what the current build reads, as the handbook states it.
Files.
Text, Markdown, PDF, Office, ODF, RTF, property lists, XML, JSON record by record, archives down to their members. The author, the dates and the sensitivity label of a document; the time and place inside a photo; ten copies of one file cost one index entry.
Mail.
A Takeout or mbox export in an evening, then the live account every hour: Sign in with Google for Gmail and Workspace, an app password elsewhere, Apple Mail and Outlook messages read as mail. Every folder the server lists, attachments as members, labels kept.
Chats and agent sessions.
Claude Code, Codex, ChatGPT and claude.ai exports, WhatsApp drops: every session rendered as it happens, so what the last agent learned is one search away for the next, and a compaction loses nothing.
Git.
Every repository on the machine, end to end: each commit rendered as correspondence with the files it changed, so a hash finds its commit and a line finds the message that introduced it.
Scans and photos.
OCR over scanned PDFs, images, HEIC and camera RAW previews, the pictures inside documents and animated images, on the machine's own GPU where it has one; a warning where a page's text order is unreliable.
Code.
C, C++, Go, Java, JavaScript, Kotlin, Python, Rust and TypeScript parsed by structure: functions, forms and identifiers as terms, with the commit and the thread that explain a line.
Runs on macOS (Apple silicon and Intel), Windows and Linux, identically. Speaks MCP to Claude Code, Claude, ChatGPT, Codex, Cursor and VS Code. People, places and dates become entities with timelines; meaning search works across languages; a rented GPU takes the heavy arithmetic for an hour on your say-so, sealed. The handbook is the full account.
What is one question your files should be able to answer today?
docenta is a commercial product of SKY, LLC, now in private preview. Tell us the question; we will tell you when the door opens, and the questions we hear shape what we build.
No account and no newsletter: you confirm your address by mail, then we write back when the door opens. Prefer mail? docenta@docenta.ai