A useful collection
A research catalog, a product list, or a table that translates your terminology into someone else’s. Share the fields you choose.
Useful work deserves to be used.Share data
Share data
and tools.
On your terms.
Let someone look up information or run a useful task. Choose what they can use, review it with your AI assistant, and share a link.
Public beta · Try it in your browser without signup or installation.
To publish your own service
- Codex or Claude Code
- Apple Silicon macOS or Linux
- A computer that stays online
Your choice of dataFederated learning & computeYour own hostingFor people and their AI assistants
Federated learning & distributed compute · BetaTrain models across organizations.
Train models across organizations.
Keep raw data in each silo.
Run multi-silo Flower (flwr) and NVIDIA FLARE (nvflare) training rounds, vertical PSI + SplitNN, or distributed map-reduce over OMOP CDM v5.4, HL7 FHIR R4, and Financial AML cohorts—without opening inbound firewall ports or exporting sensitive rows.
- Zero raw data movement & cryptographic masking.Each hospital, bank, or lab runs its own Froglet Node over local SQLite or FHIR R4 endpoints. Raw records stay local; updates use pairwise
SecAggmasking and per-sampleDP-SGD. - Locked WebAssembly & Rényi DP privacy ledgers.Every silo pins the exact same Wasm binary (
service_code_hash) forLogistic,MLP, orLoRAadapters and enforces a persistent, hash-chained Rényi DP (RDP) budget ledger with Merkle cohort commitments. - Byzantine defense, Shapley settlement & signed receipts.Silos poll outbound over HTTPS, every round emits a BIP-340 Schnorr-signed receipt, and the coordinator applies
MultiKrum/TrimmedMeanfiltering with Shapley value (φ_k) Lightning settlement gating.
Map-reduce & vertical alignment
Check cohort counts, compute global normalization across OMOP / FHIR / AML schemas, or intersect entity IDs via blinded HMAC-SHA256 PSI with zero raw rows exposed.
Flower, NVIDIA FLARE & LoRA
Dispatch Logistic, MLP, LoRA, or vertical SplitNN steps via turnkey froglet_fl adapters with pairwise SecAgg and RDP budget enforcement.
Multi-Krum & Shapley lineage
Verify every 5-artifact receipt chain offline, reject Byzantine gradients via MultiKrum, and gate Lightning HTLC settlement on positive Shapley value (φ_k > 0).
Bob has something useful. Alice needs it.A small service.
A small service.
A simple way to share it.
01 / ChoosePick what to share.
Bob asks his assistant to prepare a useful part of his work. Only the selected data goes into the shared copy.
02 / ReviewSee it work first.
Bob sees what is included, what is left out, and an example result. He approves before anything is published.
03 / ShareSend Alice a link.
Alice’s assistant reads what the service does and helps her decide whether to use it. Invitation access, when required, is sent separately.
You decide the scopeShare the useful part.
Share the useful part.
Keep the rest yours.
- Choose the data and the task.Review the exact fields or operation that others will be able to use.
- Decide who can use it.Start with public sharing, or set up invitations with expiry and request limits.
- Change your mind.Review updates before publishing them. Pause the service or revoke invitations to stop new work.
A shared result can be kept by its recipient. Stopping access prevents new work; it cannot recall a result already delivered.
For organizations
Collaboration shouldn’t start with an infrastructure project.
Agree what can be shared, who can use it, and under which conditions.
Our direction is to turn those decisions into a working service. Today, Froglet provides selected data sharing, cross-silo federated learning, invitations, execution limits, and signed records. Broader contract and SLA automation is still ahead.
A few practical questions
Understand it before you start.
Is this like a shared folder?
That is a useful starting point. Froglet shares a service: something another person’s assistant can query or run. You host it on a computer you control. It is not a hosted file-storage account, and there is no drag-and-drop web workspace today.
How do multiple organizations run federated learning without sharing raw data?
Each organization runs its own Froglet Node over its local dataset (such as an OMOP CDM, HL7 FHIR R4, or Financial AML cohort) and exposes a locked WebAssembly training service (fl.train_step.v1) and map-reduce service (compute.mapreduce_shard.v1). A coordinator using Flower, NVIDIA FLARE, or the froglet_fl SDK dispatches global weights to each silo, enforces Rényi DP privacy budgets and pairwise SecAgg masking inside Wasm, receives only updated weights and BIP-340 Schnorr-signed execution receipts, and aggregates them with FedAvg, MultiKrum, and Shapley value attribution. Read the Federated Learning & Compute guide.
Do I need to know how to code?
For a small data catalog, your assistant can help prepare it. Today you use a file-capable assistant such as Codex or Claude Code on Apple Silicon macOS or Linux. Installation and publishing still require your review. Connecting an existing system or setting up invitations needs more technical configuration.
What is private, and what becomes public?
You choose the data included in the service. Public services are accessible to others; invited services need a separate access credential. A service’s description, example, and selected package information may still be public. Keep confidential data and credentials out of that material.
Invitations currently cover the services on one Froglet host. Use separate hosts when different recipients should have different access. Read the access guide.
Can I charge for what I share?
Paid services are part of the vision, and payment integrations exist with different levels of maturity. The guided publishing journey starts free. It is not yet a ready-to-use sales and billing service. See payment integration status.
Can someone run their own program on my computer?
Yes, if you choose to offer bounded computation. Alice can supply a compatible WebAssembly program and input; Bob’s Froglet Node runs it with agreed memory, time, and execution limits. The native agent tool supports this, with an optional A2A connection between configured Nodes.
The repeatable demo uses two local Nodes and free jobs. Public setup and real payments need further qualification. This does not let arbitrary programs use your files or network. See the working flow and its limits.
Where does my service run?
On your computer or infrastructure. That machine needs to stay awake, online, and running Froglet. Managed always-on hosting is not available yet. This is a public beta; clean-machine and first-time-user qualification are still pending.
Does Froglet enforce every part of an agreement?
It can apply supported technical controls, such as selected data, invitation access, and execution limits. Signed records make an exchange inspectable. They do not establish that a result is correct or automatically enforce legal terms, data-use promises, or an uptime SLA.
What do you have
that someone else could use?
One catalog. One lookup. One useful calculation.