The Intelligent Librarian.
Nothing touches your information but the librarian.
For organizations whose information other people need, or will need: a librarian that answers for it, with sources, so people and agents get the same answer.
Every great library has equally great librarians. They do not need you to know where to look. All you need is to arrive with a half-formed query, and the librarian will quickly narrow down what it could be about, where in the collection to look for answers, and how to give you the answers for best understanding.
manba is that librarian for your own information domain, whatever it holds: books, invoices, occupations, a music collection. The library was built from your info, so the librarian knows what an invoice is here, which vendor is Acme under three names, how many records an answer rests on, and shows the sources every time. Ask again next quarter and the answer is still true, or you are told what changed.
People, AI agents and the software you already run all ask the same librarian.
Who it is for first
Your information is already being asked about.
Statistical and research bodies, public-interest publishers, trade associations, benchmark and data providers: your figures are being read, summarized and quoted by agents you do not control, and every misreading carries your name. An encounter at the border becomes a person. A preliminary monthly figure is quoted after it was revised. A share is reported as a count.
A manba librarian answers for your information, from your information, with the definitions, caveats and versions you would give if you were asked yourself. Next come the software companies whose customers’ agents are about to start asking them the same way.
Add it to your world
Which library is yours?
Bring your own information.
A statistical release, a report archive, a benchmark, an API; or a Drive or database that matters and changes. Tell the librarian what people ask of it; it builds the library and keeps it true.
Send us a dataset →
LiveA public library.
A domain the world shares: U.S. employment and careers at wywm.org, one person’s music at moodely. Use one, or propose one and steward it with us.
See the public libraries →
Build it with usBuild the librarian.
You see what this is and want to make it: engineer, researcher, domain steward, investor. The deep dive is how the library works inside, and how to reach us.
Read the deep dive →
Why
Why not an agent and a database?
Today you’d have to open the door to an LLM-based agent: give it broad access rights, have it write code and execute it in your environment, then check that it met your security and performance constraints. And at the end it still decides, each time, what an invoice is and which Acme you meant, and keeps none of it. Nothing to inspect, test or correct; when the data changes, the answers drift and nobody is told.
Intimate work, with many places to go wrong, however smart the agent is. A guess with good grammar is not information.
Two honest modes
What the data says. Or what can be derived, labelled as derived.
When the data has the answer, the librarian says what the data says. When the direct path does not serve, it analyzes: correlates, contrasts, clusters or derives a workable answer from the sources it holds, and labels it as derived, naming the technique and the evidence. What it never does is produce an answer it cannot trace.
The data says
Deterministic lookup and governed computation. Source, transforms, freshness, confidence and completeness on every result. A sum tells you how many rows contributed.
SUM AMOUNT OF INVOICES WHERE VENDOR "Acme" AND MONTH "2026-10" → 14,210.00 · 12 of 14 invoices carry an amount · epoch 41
Derived, and says so
A value the data does not state outright can still be derived for you, with the derivation disclosed: which sources, which method, why it was triggered, how confident.
DERIVE GET VENDOR FOR FILE "inv-0923.pdf" → Acme Corp · derived method: letterhead extraction + vendor list match evidence: section 1, 3 prior invoices · confidence .91
Your manba team · why this wins in year three
The librarian keeps it true. That is the part that compounds.
Every information service decays. Sources change shape, publishers revise last month’s figure, vocabularies drift, people ask questions nobody anticipated, a default quietly changes an answer. An agent and a database start from zero every time; nothing they learn is kept.
Your manba team is a set of agent roles resident in the library. They watch the sources on a heartbeat, notice what changed and what people asked that the library could not answer, propose the change in the same governed language you use, prove it against the library’s own benchmark, and release a new version. Every step is written to one open ledger you can query like the information itself. Ask again next quarter: the answer is still true, or you are told what changed and why.
Watches
Every source, on a heartbeat. A revision, a new release, a renamed column, a field that went empty: noticed, not discovered by a reader.
Proves
No change ships without passing the library’s benchmark, graded against the figures the publisher printed, not against itself.
Remembers
Every version stays answerable. Every change has a reason in the ledger. Fourteen months of hand-made releases at wywm.org, made resident.
The library
What the librarian does for you.
Learns your data once, and keeps it.
A model of your domain, built from whatever you have: tables, documents, spreadsheets, email, APIs. It stays, it can be opened, and it carries where every fact came from. This is the shared understanding you and the librarian answer from.
the information model
Lets everyone ask it their own way.
Conversation for people, a protocol for AI agents, REST and HTTP for the software you already run, which needs no rewriting to join. Same answers, same provenance, each scoped to what that client is entitled to know.
the desk
Keeps itself true, and tells you what changed.
A resident team watches the sources and the questions, proposes changes through the same governed language you use, proves them against the library’s own benchmark, and releases a version. What a good team does by hand for a year, made resident.
the engine
Security
Why nothing touches your information but the librarian.
The librarian takes on the four burdens the door would otherwise hand to an agent: the access rights, the code, your security and performance rules, the resources to work through the data. The intimate work, reading the records, joining them, analyzing them, happens inside the librarian, under your policy. What crosses out to any client is a governed result, scoped to what that client is entitled to know.
An outside agent never sees a table, a credential or a row it was not meant to see. It sees answers, with their provenance, in a language that will not let it ask for more than it should. The same boundary that makes the librarian trustworthy makes it safe to expose.
Interests grow
Tell it you care about invoices.
A manba library is not finished at setup. Say you now care about invoices, or contracts, or hiring, and your team mines the same material for the new interest: finds the documents, extracts what matters, extends the language, writes the playbooks, checks its work against examples you confirm, and releases a new version.
The same source can be read through many lenses over time, each added in a conversation, each with its own benchmark.
Proof
Two live libraries. The next proof, named.
wywm.org
U.S. employment and career data served to counselors, researchers and the public for over a year: 1,752 companies, 861 occupations. The library the engine was learned from. A manba public library. Ask it →
moodely
One person’s music library, 20,695 tracks, as a governed language. Ask for a mood, an
occasion or a narrative; PLAY builds the trajectory and EXPLAIN says why.
Analysis as the product, every answer labelled by where it came from: the library, the playlists, or you.
Ask it → · The idea →
Pick a report. Get a librarian for it.
USAFacts publishes the figures Americans look up most. We start with two of its ten most-read articles of the year, because each holds a figure that is easy to misquote:
- How many people work for the federal government? Monthly since 1939, from the Bureau of Labor Statistics. The latest months are preliminary and will be revised; the peak count (1990) and the peak share (1944) are different questions. The librarian says which figure is final, and tells you when a revision lands.
- What can the data tell us about unauthorized immigration? Border encounters, from U.S. Customs and Border Protection. An encounter is an event, not a person, and the categories changed in 2020 and 2023. The librarian knows what an encounter is here, and says so.
The benchmark is graded against the figures USAFacts printed, not against itself. A fresh agent, told nothing, answers with the table cited. Then any of the other eight.
Say “our private events.” Watch it adapt.
The same team, inside a company: a restaurant group’s real events and catering Drive, with 400 files of contracts, banquet event orders and menu sheets. Declare the interest and the team links each event across its documents, asks you to confirm the ones it is unsure of, extends the language and releases a version. “Which events next month have a BEO but no signed contract?” answered, with the files linked. Shown privately; the client names are real.
Three answers
Who, what, and why not the obvious thing.
Who is it for?
Organizations whose information other people need, or will need. First: statistical and research bodies, public-interest publishers, associations and data providers. Next: software companies whose customers’ agents are about to query them, and teams with a Drive or database that matters.
What does it fix?
Answers nobody can check, and agents on raw data nobody can contain. The second is a security exposure. The first is a slower one.
Why not an agent and a database?
Because the understanding has to live somewhere. wywm.org took fourteen months and 155 hand-made releases to keep true. That is the work the librarian does by itself.
Start here
Send us one dataset. Get three answers back.
Point us at a dataset you publish or depend on, and three questions people get wrong about it. We will build a librarian for it and send you its answers, with sources, and with what it could not know. Or ask for a walkthrough of wywm.org.
Received. A person will reply, with the answers or with questions.
Meaning has an owner. Make it yours.
Every manba library declares what it is for: its scope, the facts its sources cannot know, its methodology, the questions that matter. That declaration is what separates a library from a search box, and it is what lets you build one only you could have built.