Hardware & RF · Daniel Tone
Software can claim transport independence. Radios settle it.
Every architectural claim on this site ends at the same place: a real antenna, in real weather, at real distance. That is Dan's half of FrogNet — the physical layer, the radios, the New York nodes, and the bring-up work that turns a design decision into something that either holds a link or doesn't.
Hardware — what must be hardware, and what only must be software.
Co-founder
Daniel Tone
VP of Engineering · hardware & RF
- Owns
- Physical layer, radios, antennas, bearer bring-up
- Nodes
- The New York side of the production mesh
- Field
- HaLow, LoRa, Meshtastic bridges, ESP32 sensor and actuator boards
- Also
- Cross-platform bring-up
FrogNet is two people, and the division is clean: I write the software and the algorithms; Dan makes them survive contact with the physical world.
That split matters more than it sounds. A protocol that adapts to link characteristics rather than to a medium is a claim you can make on paper and be wrong about for years without noticing — right up until someone puts it on a 900 MHz radio in a moving vehicle and watches what actually happens. Dan is the one who finds out.
The heterogeneous path the Communicator demo runs over — laptop to Wi-Fi to a Seattle node, across WireGuard through the broker, into a New York node, out over 900 MHz to a second New York node — only exists because someone built the far half of it. The adaptation responding to bandwidth, latency, jitter, and loss rather than to any particular bearer is the architecture's claim. Watching it hold across that path is Dan's evidence.
The credit line is worth stating plainly, because I am careful about authorship elsewhere: the algorithms are mine. The hardware, the RF, the New York nodes, and the cross-platform bring-up are Dan's.
What held up, and where.
Not a lab bench. These are the runs that turned architectural claims into observed behaviour.
Mobile · Queens, New York
A trunk-mounted node, driven through a dense city
A HaLow node in the trunk of a moving vehicle, driven through Queens — holding the link continuously across multipath, building occlusions, and the metal enclosure of the car itself. No observed drops. IEEE 802.11ah on the sub-gigahertz ISM band presents to Linux as an ordinary Wi-Fi interface, so FrogNet rode it with no protocol translation at all.
Proves · continuous mesh participation while mobile
Closed loop · 2,400 miles
A button in New York, a database in Seattle, an actuator on the desk
The first independent end-to-end demonstration: a Flask UI on a New York Wi-Fi network, an ESP32 sensor and an ESP32 actuator on a Manhattan desk, all coordinating through FrogNet Memory, hosted in Seattle — the whole loop over semantic-compressed HTTP, with no message ever addressed from one device to another.
Proves · sensors and actuators as shared memory, wide-area
Bridged bearers
Sensor data arriving over radios narrower than IP
Meshtastic and generic LoRa bridges writing readings straight into FrogNet Memory, validated end to end. A sensor on a ridge with no IP infrastructure of its own becomes pond-wide current state the instant it lands — visible to every node and every AI at once.
Proves · the bearer is irrelevant above the transport line
The flagship path
The far end of the transcontinental call
The New York half of the Communicator run: WireGuard in from the broker, then 900 MHz out to a second node. The takeaway was never the resolution — it was that the conversation never stopped while the link was deliberately starved and recovered.
Proves · transport independence, shown rather than claimed
Adaptation that responds to a medium is a driver. Adaptation that responds to characteristics is an architecture. You cannot tell which one you built without the radios.
— the reason this page existsWorth noticing what that story actually is. Dan didn't apply for anything — he had the RF background, saw a problem worth solving, and the hardware and RF work became his because he did it. That is how every area here is meant to change hands, and there is no reason it can't happen again with somebody reading this. How responsibility works →
The bearers, and what each is actually for.
None of the application code knows the difference. None of the pond and chorus logic changes. New bearers plug in below the line; the architecture above it does not move.
Ethernet · 5 GHz Wi-Fi
Short range, high bandwidth. The ordinary case, and the baseline everything else is measured against.
900 MHz HaLow
802.11ah on the sub-gigahertz ISM band — licence-free, roughly two miles in clear air, full IP. Presents as a standard Wi-Fi interface to Linux.
WireGuard over the internet
Where a broker provides a path between sites that cannot otherwise reach each other. The encryption, and the only public surface.
Meshtastic · LoRa bridges
For sensor data arriving over radios too narrow for IP to ride directly. The bridge writes into the shared store; the reader never knows.
For the guild · open problem
Bring a bearer we haven't put it on.
Transport independence is only as proven as the number of transports it has actually survived. If you work with a radio, a link, or a physical layer we have not tried — that is a contribution, and the result is interesting whether it holds or breaks. Dan lives in the hardware forum. Ask for a licence → — and if you want to work on this, say so and say which. The questions are optional and the licence does not depend on them.
The physical half
Radios, antennas, and what held up at range.
If bearers and field deployments are your half of the world, that is a seat at the bench — and it is the one place where the architecture gets told it is wrong by something other than an argument.
The bearer is the part we have least of
We need: hardware and RF people, and anyone with a platform we do not own.
FrogNet runs over whatever you have — cable, Wi-Fi, radio, tunnel — and the bearers it has been proven on are the ones two people could get their hands on. New bearer work, antenna work, and measurement on hardware we do not own are all contributions that need no engine access.
One in particular is worth naming, because nobody has the table. There is no per-participant term at the edges of a call — that is architectural — so how many participants a call supports is a property of the box holding mediahost.frognet: cores, encoders, hardware acceleration, and whether it is doing five other jobs. Measure the ceiling on your hardware and report it. No engine access, no Python. The capability probe already emits the fields that explain the spread, so a result arrives with its own explanation attached.
How you would know it worked. A participants-against-platform table, from a Pi to a dedicated encoder box, each row carrying the probe output that explains it.
None of these is hidden behind a fallback. Each one fails visibly, says what it is, and is written down — which is the only honest way to ship something unfinished. Hardware & RF → · Take the free licence →