Inspiration
The most useful things an AI agent could do are the things we can't let it see: your medical records, your company's accounts, your private code.

So today you pick. Upload everything to someone else's cloud and hope, or keep it safe and do without the help. Either way, the work that matters most is the work the best agents never touch.
There is a second problem, and it surprised us. You can't tell which agent is good. We measured it. One helper, the same model (NVIDIA Nemotron), the same sixteen questions, three versions: 9, 13 and 14 right. Only the setup around the model changed: its instructions and its checks. From the outside you can't see a setup, so you choose blind, and to try one you have to hand over your data.

What it does
agent-bridge flips the direction. In Pull mode the expert's setup (we call it a helper) comes to your machine. Its stepses your own Nebius key, and it only sees what you open for it.

- You tick the shelves. My memory is a library of about 140 notes on shelves. Before you hire a helper you tick the shelves it may read. A secret note stays home even if you tick its shelf.
- A gatekeeper, not a search box. The helper can't search your notes. The memory brokerhands it only what is on the shelvead, and plants decoys that only anout-of-bounds query could reach. - **Nemotron on Nebius Token Factorour key, which the helper neversees.
- You see what left. Every answer lists the notes it used, which files were sent to the model and how much text. Emails, phone numbers and account numbers are swapped for placeholders before the model call and put back in the answer on your machine.
- Independent scores, not stars. A certifier runs every helper on the same questions withe same Nemotron model and ranks tanks above one with 9.
- A cheat is stopped in a sealed room. Code a helper brings runs with no network, a read-only disk and no extra powers. In our tests a helper that tried to send my data out was killed with nothing delivered, and so was one t never shared. Hard evidence in alive run takes the helper off the market at once.
- Anyone can sell a setup. In the studio you arrange the steps, set a price and publish. Before it goes on sale the platform checksead private data. Buyers pay peruse, with test money in this demo, and the author never sees their notes.

Imagine a helper for your tax papere that reads your whole codebase.The same room and the same rules, whatever is on your shelves.
How we built it
- Workflow engine (
flowd). A model call, code, gate, return).Every step goes into a hash-chained journal, so what happened can't be quietly rewritten. - Memory broker. The only way a helper reads notes; stages records from ticked shelves, counts reads, serves decoys to off-lens queries.
- Result guard. Bounds the only exit: schema, size, decoy values, verbatim copies of your records. Schemas that allow unbounded text are rejected when a helper is published.
- Sealed cell. Code steps run inetwork, a read-only filesystem andno capabilities, with a sensor that records attempts to use the network, start a process or write outside scratch space.
- Inference. NVIDIA Nemotron (
-30B-A3B) on Nebius Token Factory,with bounded retries, and a masker that replaces emails, phone numbers and account numbers with placeholders before the call. - Web (Tavily). Lookups are allowed only in steps that can't read private data; the platform checks that before a helper can be published.
- Certifier. Same questions, same model, Wilson lower-bound scoring (a helper that passes 3 of 3 scores lower than one that passes 28 of 30, on purpose), signed results and a security trap.
- Console. A pixel-art market floor where each chat is a stall, a memory library, and a studio where you draw a helper as a pictur
Challenges
- Making "the helper can't send anything out" a property of the runtime, not a promise. We went through two designs before settling on a closed language plus a no-network container.
- **Being honest about the model sees read what a helper's steps sendit. We can't change that, so we show exactly what was sent, mask what we can, and say what we can't.
- Nemotron's empty replies. Nemotron reasons in a hidden field and can use its whole token budget before it answers, which looked like empty replies. We found it by reading
finish_reason, turned thinking off (complete answers in about 40 tokens instead of 300 to 1,200), retried, and made the certifier ignore infrastructure failures. - Honest evaluation. Our first marks. A harder suite spread themout.
Accomplishments
- A measurable result: three versions of one helper, the same Nemotron, the same sixteen questions, scoring 9, 13 and 14. The ranking shows setups, not models. - A helper with real malicious coded one that reaches for a bait noteis caught by behaviour.
- About 660 automated tests, and a README quickstart.
What we learned
A marketplace for agents only works if quality is measured independently, and privacy only works if it is a property of where the code runs and not a promise. The safe design is the boring one: a small language the platform can reason about beats a flexible one it can't. ## What's next
Hosted mode on someone else's machine, sealed enclaves, multi-device memory, and verified PayPal sandbox payments.
Honest limits
- The model service reads what a helper's steps send it (minus placeholders). Information already shown to a helper can't be taken back.
- Detection is evidence, not proof. A slow leak through allowed output may not trip anything.
- The certifier suite is small. Scoite", never "safe".
- Pull mode has no fair exchange: the buyer's machine could keep an output without paying. Payments limit the loss; they don't remove it.
- Payments in the demo are test money only.
Built With
- ai-agents
- cryptography
- css
- docker
- ed25519
- html5
- javascript
- json-schema
- nebius-token-factory
- nvidia-nemotron
- privacy
- python
- pyyaml
- sandbox
- tavily