Runs in simulation today, on real terrain. Nothing has been flown.
Search drones you can talk to.
Reports come in from people in different places and in different words. Meshfare weighs them, proposes what the swarm should do, and lets ordinary code check every instruction before a drone moves.
A language model is good at reading what people write and poor at being trusted with aircraft. So it is given one job: turn a report into a short plan from a fixed menu of nine instructions. Everything after that is ordinary, tested code.
A report arrives, with its source
From a control room, a ranger, a ground team's phone, or typed by the coordinator. Who it is from is fixed by the link it came through, never by what the sender claims.
A plan is proposed
A model, or keyword rules when no model is available, reads the report beside a brief of the search and proposes instructions from the menu. It sees nothing the search itself does not know.
Every instruction is checked
Shape, place and possibility are validated in code. Limits such as how many drones must stay on the general search are enforced there, not left to a prompt.
Out of proportion? It waits for a person
A plan that would move most of the swarm on weak evidence is held, with the reasons, until someone approves or dismisses it.
control room · demonstration
What it does
One picture of the search, for everyone who needs it.
An operations console on the real ground
The search on a satellite map or the 3D terrain of the place: a shared probability map that every look updates, the drones with their state and battery, the reports and the plans they led to, and a log of everything.
The fleet on one page
Every aircraft with its battery, link and position. Recall or release one, several or all.
Reports in words, or on the map
Type what a coordinator would say, or circle an area and write about it. A vague place, "by the lake" on a map with three, is asked about before anything is planned.
A search anywhere on Earth
Pick a point. The app fetches its heights, land cover and place names from open data and builds the map to search.
Fits what you already run
Other systems read the search as GeoJSON and send reports as JSON with a latitude and longitude.
Field teams
The people on the hill see the same search.
A team on foot opens a link on the phone they already carry. No account, no app store.
Searched ground, never "cleared". How thoroughly each piece of ground has been looked at, in three steps, because a search can miss.
One large button to report. With a circled area or the phone's position. It arrives in the control room under the team's name.
Works with a poor connection. A report written without signal is kept and sent when there is one.
Shown only what a team needs. Never the probability map, the plans or anyone else's reports.
Ground team 1● live
looked at lightlysearched wellsearched thoroughly
Report something
Built to be trusted
Limits live in code, where they can be tested.
01
A fixed menu of nine
A model can only produce instructions from the menu. It cannot write code, name a waypoint outside the map, or reach an aircraft. With no model at all, the search still runs.
02
A record that shows tampering
Every command, report and find is written as it happens, each entry carrying the hash of the one before, with who was signed in when it was done.
started→report→plan→accepted→found
03
Your network, encrypted
Devices connect over HTTPS with a certificate made on the coordinator's own computer. No outside service sits in the middle, and the control room stays on that computer.
Evidence so far
Measured, with the caveats attached.
Everything below was measured in simulation, on 60 real lost-person probability maps or on small test sets written by the author.
20of 20
Everyday wording understood
Twenty reports in loose, everyday words: Claude Opus 5.5 read all 20 correctly. The keyword rules, with no model, read 8.
+9points
Found more often with a model planning
One true and one false report per search, 65 searches of 30 minutes: found in 35% with Opus planning, 26% with the rules, 8% with reports ignored. 95% interval for Opus against the rules: +1 to +17 points.
~2times
Camera strip against hovering
Drones that search with a camera strip while moving find roughly twice the people of drones that stop and hover over each cell.
What this is not. No aircraft has flown under Meshfare. The test sets are small and the camera model's values are placeholders. It is a research tool and a demonstration, not something to run a real search with today.
Early access
See a search run on your own ground.
Meshfare is not yet generally available. It is being built with search-and-rescue teams, researchers and drone operators who want to try it, in simulation, on terrain they know.