01
Consultant frameworks & playbooks
The method you charge for. Diagnostics, pricing models, negotiation scripts, the judgement calls that took a career to learn.
Your knowledge earns.
Nobody downloads it.
Point KBGate at what you know and set a price per answer. AI agents pay per question and never download your raw files.
Free to list. You are charged out of what you earn, never before.
One request
Per-answer billing
Agent asks
KBGATE
Answer leaves+$0.05
Your filesnever move
Each question is a sale of one answer, not of your library.
What you gate
If people already pay you to answer it, it belongs behind a gate.
The test is simple: knowledge specific enough that a general-purpose AI gets it wrong, and valuable enough that being wrong costs someone money.
02
Proprietary research & models
Datasets you maintain, benchmarks you run, survey results, forecasts. Work nobody else has and everybody else needs.
03
Private codebases & domain docs
Internal runbooks, architecture decisions, integration quirks, the regulatory detail that only matters in your industry.
How it works
Three steps, and the meter starts.
Nothing to build, no API to design, no contract to negotiate. You upload, you price it, and the AI tools people already use can start paying to ask.
01 UPLOAD
Point it at what you know
Drop in a folder of documents. KBGate reads it, keeps it private, and gives you one link that answers questions about it.
02 PRICE
Set a price per answer
From a cent up, and you can change it whenever you like. The price is also your protection: asking your way through everything costs more than licensing it.
03 EARN
AI agents pay to ask
Every answer that goes out is a charge and a line in your ledger. Blocked questions earn nothing and reveal nothing. Your files never move.
Hover a step for detail. The gate does the same work on every question.
Planned shape of the upload tool — in build
$ npx kbgate upload ./field-notes read 148 documents · 2.3 MB · kept private library kb_7f3a2c price $0.05 per answer link https://kbgate.com/q/kb_7f3a2c Share the link. You are paid for every answer it gives.
For technical evaluators
Adversarial enforcement & harvest defense
How the runtime defends proprietary IP against systematic agentic reconstruction and latent-space distillation. The threat is not an intruder. It is a paying customer with a valid key, asking well-formed questions at machine rate, whose goal is to stop needing you.
- Adversary
- Authenticated caller in good standing. Can rephrase without limit, vary embeddings, distribute across sessions, and wait. Any defence that assumes clumsy probing is worthless.
- Objective
- Reconstruct enough of the corpus to stop paying, or distil the answers into a local model that answers for free.
- Security goal
- Not prevention of information transfer — an answer that carries no information is not worth buying. The goal is to make full reconstruction cost more than a licence, and to make theft provable after the fact.
- Out of scope
- A human reading answers and remembering them. Nothing here defends against that, and no honest system claims otherwise.
- Status
- Design specification. Not deployed. Written down so it can be argued with before it is built.
01
Zero-chunk exposure
Synthetic egress only
- Raw vector chunks, verbatim text spans and source embeddings are never serialised to the egress layer.
- The response schema is a closed type. There is no field capable of carrying a span, so withholding source is a property of the interface, not a filter applied to it — and a filter is what gets talked past.
- Every output is either a deterministic tool result or a synthesis composed inside the runtime.
LIMITA useful answer is information. This bounds the rate of transfer, not the total. Pillars 02 and 03 bound the total.
02
Stateful latent coverage tracking
Durable Objects · strongly consistent
- One durable object per library holds the coverage map. Single-writer, so accounting is strongly consistent and cannot be dodged by spraying requests across regions or connections.
- Coverage is occupancy over cells of the embedding space, not a request count. Rephrasing the same question lands in the same cell and buys nothing.
- Systematic tiling and vector triangulation of the space are visible as a shape in the coverage map long before the space is covered. On detection the runtime degrades resolution and answer fidelity.
LIMITCoverage is tracked per identity. Spreading a harvest across many accounts defeats it, so this leans on account verification, which is a separate problem.
03
Economic & entropy asymmetry
Harvest deterrence
- Price is a function of marginal novelty, not volume. A genuinely new question stays at list price no matter how many came before it, so heavy legitimate use is never punished.
- Questions landing in already-covered cells escalate steeply. Probing the same latent boundary burns quota superlinearly.
- The escalation curve is set so that the cost of full coverage is a large multiple of the licence price.
LIMITThat multiple is a parameter, not a theorem. It is only as sound as the owner's valuation of the corpus, which is why the owner sets it.
04
Semantic watermarking & provenance
Attribution after the fact
- A keyed, distortion-free watermark is embedded during synthesis. It is carried in sampling choices, so output quality is unchanged and there is no visible artefact to strip.
- The watermark is radioactive: a model fine-tuned or distilled on leaked outputs carries it into its own generations.
- The key is cryptographically bound to the issuing API key. Detection is a hypothesis test whose confidence rises with the number of sampled outputs — enough text, and attribution is not seriously disputable.
LIMITAttribution is statistical, not a cryptographic proof. Heavy paraphrase and short outputs both degrade detectability. This deters and evidences; it does not prevent.
What you earn
Set your price. See the month.
Small payments only feel small one at a time. An AI agent does not ask once; it asks on every task, every draft, every ticket, all day.
$0.05
$0.01$1.00
250
102,500
You keep
$337
per month
$0.05 per answer × 250 a day
gross $375 · platform fee 10% −$38
Planned fee. Nothing is charged until an answer sells, and the demo below bills at the price you just set.
What never crosses the gate
You are selling answers, not access.
This is the part that makes it safe to say yes. Every rule below is checked before anything is sent back, and you can watch each one work in the demo.
HELD BACK
Your files
Answers are written behind the gate. The original documents are never sent, downloaded, or attached. A buyer can quote you; they cannot copy you.
HELD BACK
The whole library
No bulk download, no table of contents, no "just give me all of it as a spreadsheet". Handing everything over is not something the gate can do, so no clever wording makes it happen.
HELD BACK
Everyone else
One library per question. Nobody else's material is ever in range, and yours never turns up in someone else's results.
Interactive demo
[ Simulation • Demo environment ] Runs in your browser at the price you set above. No real money moves.
KBGATE demo
YOUR LIBRARY in range
ANOTHER OWNER out of range
Two libraries. Only one is ever in range.
PolicyWhat sells, what stops
| # | request | decision | you earn | rule |
|---|
Before you start
The questions owners actually ask.
What kind of knowledge is worth gating?
Anything specific enough that a general-purpose AI gets it wrong. A consultant's playbook, a dataset you maintain, an internal runbook, a decade of field notes, the pricing quirks of one industry. A good test: if people already pay you to answer the same question over and over, that is a library.
Do I hand over my files?
No. You keep the originals, and KBGate answers questions from them without ever sending them out. The only thing that leaves is the answer you priced, and you can pull the material at any time.
How do I get paid?
Every answer that goes out is a charge to whoever asked and a line in your ledger. The balance builds up per question and pays out on a schedule you choose. The planned platform fee is 10% of what you earn, with no listing fee and no monthly minimum.
Can't someone just ask a thousand questions and rebuild it?
That is the real risk, and the price is the defence. Every question costs, so working through your whole library has a bill attached. Each buyer also gets a spending limit, and if someone starts asking the same thing in slightly different ways or marching through a topic in order, the price climbs or the questions stop.
The economics are the point: rebuilding your library one answer at a time should cost more than simply paying you for the answers. If you want the mechanism rather than the summary, the enforcement spec sets out the threat model and where each defence stops working.
What does it cost me to run?
Nothing to set up and nothing to keep listed. You are charged out of what you earn, so a library nobody asks about costs you nothing.
Can I change the price, or shut it off?
The price is a single field; change it whenever you like and the next question bills at the new rate. Cutting off a buyer is one switch, and taking the whole thing down is another. The material stays yours throughout.
Is any of this live yet?
Not yet, and we would rather say so here than bury it. The gate is in build, the upload tool shown above is the planned shape, and the demo on this page is a simulation running in your browser — the rules and decisions are real, the traffic and the money are not.
That is why the button says Request Access rather than Sign Up. Access goes out in order, and the first cohort gets a library, a price, and a working link before anyone else.
Request access
You already have the knowledge. Put a price on the answer.
Join the owner cohort. You will get a private library, a price per answer, and a link that AI agents pay to ask.
Free to list. Charged only out of what you earn. No spam, one email when your gate is ready.