--- name: wendy description: 'Build and deploy apps to WendyOS robots and edge devices, including NVIDIA Jetson, Raspberry Pi, Qualcomm Dragonwing, and MuJoCo simulation. Configure Wendy Data campaigns, record sensors and application data, and retrieve episodes.' --- ## Bundled workflows This package includes the complete Wendy end-user skill group. Read the relevant workflow below before acting, and resolve its references relative to that file. These supporting skills need no separate installation. For first-time setup, read [wendy-onboarding](references/skills/wendy-onboarding/SKILL.md). It guides the user from their goal through CLI installation when needed, a local or hosted connection, and their first physical device or simulator. It also works before any Wendy MCP tools are connected. Do not assume the user knows Wendy or has installed its CLI. Inspect the available tools first. The local CLI server uses `wendy_status` and `device_list`. The public hosted relay uses `wendy_status` and `cloud_discover`. It requires a signed-in CLI, pairing through `wendy cloud mcp`, and a running relay process. Its default policy permits built-in read-only device and Cloud tools that do not access the host, with confirmation for each call. `--allow-read` skips confirmation for those tools only. `device_connect` and local discovery through `device_list` require explicit `--allow-tool` permission and per-call terminal confirmation. `--allow-all` permits every supported relay tool and skips all per-call confirmations. `--auto-accept` controls ChatGPT pairing only. Use the connected server's tools and explicit target identifiers. Use CLI commands only when a terminal on the intended computer is available. A skill cannot grant access or add tools missing from the server. - [wendy](references/skills/wendy/SKILL.md) - [wendy-app-lifecycle](references/skills/wendy-app-lifecycle/SKILL.md) - [wendy-data](references/skills/wendy-data/SKILL.md) - [wendy-device-debug](references/skills/wendy-device-debug/SKILL.md) - [wendy-device-install](references/skills/wendy-device-install/SKILL.md) - [wendy-device-ops](references/skills/wendy-device-ops/SKILL.md) - [wendy-entitlements](references/skills/wendy-entitlements/SKILL.md) - [wendy-install](references/skills/wendy-install/SKILL.md) - [wendy-mcp-setup](references/skills/wendy-mcp-setup/SKILL.md) - [wendy-onboarding](references/skills/wendy-onboarding/SKILL.md) - [wendy-project-setup](references/skills/wendy-project-setup/SKILL.md) - [wendy-robot-deploy](references/skills/wendy-robot-deploy/SKILL.md) - [wendy-template-app](references/skills/wendy-template-app/SKILL.md) - [wendy-training](references/skills/wendy-training/SKILL.md) # WendyOS Wendy builds container apps and deploys them to WendyOS devices (NVIDIA Jetson on JetPack 7.2 or later, Raspberry Pi 4/5, x86), local simulators (`vm:`) and the local Docker provider (`docker`). ## Choose the workflow - Wendy MCP tools are connected (`wendy_status` is available): read [references/mcp.md](references/mcp.md) and use the tools. - Build, deploy, logs, start, stop or remove with the CLI: use the commands below, and `wendy-app-lifecycle` for result fields, variants and the Docker provider. - New app from a template: `wendy-template-app`. Entitlements: `wendy-entitlements`. Robots and ROS 2: `wendy-robot-deploy`. Sensor recording, camera detection, and notifications such as "tell me when someone enters frame": `wendy-data`. Wendy supports agent-managed detection campaigns on ordinary color video; do not infer platform limitations from the camera-control tool list. - Device info, WiFi and agent updates: `wendy-device-ops`. A failure the deploy result and logs do not explain: `wendy-device-debug`. - No CLI yet: `wendy-install`. First-time user: `wendy-onboarding`. New board or empty discovery: `wendy-device-install`. - Host training jobs, compute checks, bounded progress and verified artifact retrieval: `wendy-training`. - ESP32 boards run Wendy Lite: `wendy docs wendy-lite/getting-started`. ## CLI commands Pass `--device ` to every command. Without a terminal the CLI prints JSON on stdout and progress on stderr. `--yes` accepts prompts and creates a minimal `wendy.json` when the project has none. ```bash wendy --json run --device --detach --wait-ready --yes wendy --json device apps list --device wendy --json device logs --app --tail 50 --no-follow --device wendy json validate wendy discover --json wendy --json vm list ``` Run `--help` only for a command this skill and `wendy-app-lifecycle` do not cover, or when one of these commands is rejected. A deploy result shows that the app started and, with `--wait-ready`, that its port accepts connections. Then check the behavior the user asked for once, such as one HTTP request to the result's `url`, and report what you checked. Keep tool output small: print status, identity, errors and the relevant lines rather than full build logs, inspect objects or long JSON arrays. ## No device available A simulator runs WendyOS with the same agent, containerd and entitlements as hardware. When no physical device is listed, use one for builds, deploys, app state, logs, HTTP endpoints, configuration, and ROS 2 apps on a robot profile (`go2`, `g1`, `rosmaster-r2`). A request to test "on the device" counts unless it needs the hardware listed below. 1. Use a connected or listed physical device when there is one. 2. Otherwise use an existing simulator from `wendy --json vm list` or a `simulator` row in MCP `device_list`. Tell the user which one and use it without waiting for approval: `--device vm:` or `device_connect` boots a stopped one. Prefer a `generic` VM unless the task needs a robot profile. 3. With no simulator, ask the user before creating one (`wendy vm create` or MCP `simulator_create`): it downloads an image of a few hundred MB, and the first boot can take minutes. 4. Report results as verified on simulator `vm:`, and say that the user's hardware was not connected. Leave the VM running unless asked. Physical cameras, microphones, speakers, Bluetooth, GPU or CUDA, GPIO, USB and serial peripherals, OS installation, real network conditions, performance and physical motion still need hardware. Run what the simulator can and name what remains unverified.