Bench measurementIllustrative image
The fullstack physical AI platform
Simulate on your laptop, ship the same project to real hardware, and run it in the field, from your terminal or coding agent.
Run your first projectTeams working with Wendy
Clean Core Research
The full stack
Wendy is the whole stack, already assembled
Wendy brings the operating system, logging, networking, AI inference, and app deployment together, so your team builds the app on top.
Illustration
- Operating systemWendyOS is a Linux distribution built for Jetson and Raspberry Pi, with A/B updates that roll back on their own.
wendy os updateShipped - LoggingLogs, metrics, and traces from every app, through the OpenTelemetry endpoint built into the Wendy Agent.
wendy device logsShipped - NetworkingUSB-C with no network setup, Wi-Fi from the CLI, Wendy Cloud in beta, and Wendy Net in preview so devices can talk to each other.
wendy discoverShipped - AI inferenceRun models on the device with GPU access for NVIDIA CUDA, AMD ROCm, and Qualcomm Adreno.
{ "type": "gpu" }Shipped - App deploymentwendy run builds your app for the target, sends only the chunks the device is missing, and restarts it.
wendy run --watchShipped
Runs on
Any arm64 or x86_64 Linux
In the field and on the bench
What teams have built and tested with Wendy
Autonomous farm and lawncare equipmentIllustrative image
$4M
In product development costs a robotics CEO says Wendy would have saved
CEO & Co-Founder, pilot feedback
Europe Embodied hackathon
Refinery production improvementIllustrative image
Fortune 50
Oil company testing WendyOS on a Jetson to tune refinery control loops
Automation Manager, pilot feedback
SimulationIllustrative image
0 robots
Needed on your desk to start, with a simulated Go2 or G1 in a WendyOS VM
From simulation to real devicesThe Wendy platform
One workflow from simulation to the fleet
Wendy handles target builds, deployment, hardware permissions, and logs so you spend less time maintaining device scripts.
Start in simulation, with no hardware on the desk
Your coding agent tests on a simulated robot in a WendyOS VM, using the same project and wendy run workflow you will use on the device.
Simulated Go2Preview
$ wendy vm create go2-sim --profile go2
$ wendy vm robot start go2-sim
$ wendy run --device vm:go2-sim
Application patrol started.
One workflow for simulators and robots
Keep the same project and wendy run workflow as you move to a robot, and run the documented checks before hardware.
What you can simulate today
- Unitree Go2Preview
- Unitree G1Preview
- Any Wendy appShipped
Proven at a hackathon
A three-person team took an autonomous inspection stack from simulation to a physical Jetson.
Ship the same project to the real machine
Deploy with wendy run over USB-C, the local network, or Wendy Cloud. Wendy builds for your target, sends only the chunks the device is missing, and restarts the app.
USB-C to a JetsonShipped
$ wendy run --device wendyos-zestful-stork.local
Application patrol started.
patrol: navigating 4 waypoints
Spend less time on device scripts
Wendy handles target builds, deployment, hardware permissions, and logs.
“If we found Wendy two years ago, we would have saved $4M in product development costs.”
Declare hardware access once
List GPU, camera, networking, and storage access in wendy.json, so every app gets only what it declares.
// wendy.json
{
"appId": "dgx-spark-inference",
"platform": "linux",
"entitlements": [
{ "type": "network", "mode": "host" },
{ "type": "gpu" }
]
}Your coding agent runs every stage
Add the Wendy plugin to Claude Code or Codex, or connect any MCP client. The agent builds, deploys, reads logs, and tests each edit on the device.
Coding agent sessionShipped
Write a patrol app for the Go2. Test it in simulation, then ship it to the real robot.
● Edit main.py
● Run wendy run --device vm:go2-sim
Application patrol started.
● Check simulation logs · no collisions
● Ship wendy run --device wendyos-zestful-stork.local · real robot
Application patrol started.
Guardrails the agent works inside
Provisioned devices accept the agent only over mutual TLS, and every MCP tool is labeled by what it can change.
Read-only
Mutating
Destructive
mTLS · device pinning · entitlements
Tested against real control loops
A Fortune 50 oil company is testing WendyOS for refinery production improvement.
“I’m testing WendyOS on a Jetson to enable in-field AI to tune old, dumb control loops. I call it Automated Dynamic Local Loop Control.”
Wendy for developers
Keep your Dockerfiles and ROS 2 code
Wendy builds for your target, transfers the app, starts it, and brings back logs. Write in the language your robot already uses.
Read the docs- ROS 2
- Foxglove
- MuJoCo
- Docker Compose
- OpenCV
- YOLO
- CUDA
- ROCm
Hardware
Deploy to the hardware you already have
Install WendyOS on a supported board, add Wendy Agent to the Linux machine you already run, or flash Wendy firmware onto an Espressif ESP32.
Security with a public status for every control
Once a device is provisioned, every connection requires mutual TLS. The docs list which controls have shipped and which have not.
Wendy · Security controlsStatus
- mTLSMutual TLSPer-device certificates on every connection once a device is provisionedShipped
- ML-DSAPost-quantum authenticationIn use today for supported device and operator identitiesShipped
- A/BOS updates that roll backA failed boot or post-update check returns the device to the previous versionShipped
- SBOMVerifiable releasesEvery release ships with an SBOM and SLSA build provenanceShipped
- SOC 2SOC 2Audit work is underway, and Wendy is not yet certifiedIn progress
Get started
Try Wendy with one app and one device
Deploy an app, change it, and inspect the result. Start with a local simulator if you do not have hardware yet.
Run your first project