# RuView plugin: set up and run WiFi sensing (ruvnet/RuView) — skill card for AI agents

> AgentGid skill card. Checked on 2026-11-12 against commit ce8d296 (version 0.3.0).
> Every fact comes from the pack's own files at that commit. Numbers marked "as reported" are the author's.
> Human page: https://agentgid.com/skills/ruvnet/ruview/

> Note: the pack has changed since this check; re-read its files before relying on details.

## Explain it to a non-programmer
A guide for your AI assistant to the RuView project. RuView uses WiFi signals and cheap ESP32 boards to sense people in a room: presence, breathing, heart rate and body pose. Your assistant walks you through the demo, flashing and setting up the boards, running the sensing apps, training models and checking a build.

## Who it is for
- You are new to RuView and want the shortest path to a working demo
- You have an ESP32-S3 or ESP32-C6 board and want it flashed, set up and streaming
- You want to run presence, breathing, heart rate, pose, sleep or search-and-rescue sensing with RuView
- You train or check RuView models and builds
- Not for: You do not work with the RuView code: the commands assume its folder layout
- Not for: You need a certified medical or safety system: the project calls its apps research prototypes

## What it is
- Kind: command, knowledge, scripts
- License: MIT
- Repository: https://github.com/ruvnet/RuView

## What it does
- Start: pick a path by the hardware you have: a Docker demo with simulated data, a build from source, or a live ESP32 board
- Hardware: build and flash the ESP32 firmware, write WiFi and network settings to the board, check that data arrives
- Configure: firmware variants, board settings, sensing server options, several boards working together
- Apps: presence, breathing and heart rate, activity and falls, 17-point body pose, sleep, room mapping, search and rescue (MAT)
- Train: five model tracks, from no-camera pose to camera-supervised pose, with optional paid cloud GPUs and publishing to Hugging Face
- Verify: run the test suite, a fixed-input proof check and build a signed evidence bundle before merging
- Radar: 60 GHz and 24 GHz radar modules, alone or combined with WiFi

## Skills
| Skill | Call | Changes code? | In plain words |
|---|---|---|---|
| ruview-quickstart | /ruview-start [docker|build|hardware], or "how do I get started with RuView" | unclear | Gets a newcomer from nothing to a running sensing dashboard |
| ruview-hardware-setup | /ruview-flash [8mb|4mb] [COM port], or "flash my ESP32 node" | yes | Builds and flashes the board firmware and checks the board sends data |
| ruview-configure | /ruview-provision --port ... --ssid ... --password ... --target-ip ..., or "tune my RuView deployment" | yes | Changes how a deployed system behaves without changing the program |
| ruview-applications | /ruview-app [presence|vitals|pose|sleep|environment|mat|pointcloud] | unclear | Runs one of the sensing apps and helps pick the right one |
| ruview-model-training | /ruview-train [camera-free|camera-supervised|embeddings|domain-gen|snn|gpu] [--epochs N] | yes | Trains, tests and publishes RuView models |
| ruview-advanced-sensing | /ruview-advanced [multistatic|cross-viewpoint|tomography|field-model|intention|adversarial|security] | unclear | Guides research features that combine several boards and links |
| ruview-cli-api | plain language ("script a survivor scan", "serve RuView over HTTP") | unclear | Explains the RuView command-line tool, web service and browser build |
| ruview-mmwave | plain language ("set up the MR60BHA2 radar") | unclear | Sets up 60 GHz and 24 GHz radar modules, alone or with WiFi |
| ruview-verify | /ruview-verify [tests|proof|bundle|all] | unclear | Runs the tests, a fixed-input proof check and builds an evidence bundle before merging |
| ruview-rvagent | plain language; its tools come from the rvAgent server | unclear | Notes for wiring agents to live presence and vitals events |

### ruview-quickstart
- Asks for: Which hardware you have: none, a computer to build on, or an ESP32-S3 / C6 board
- Gives: Tier 0: a Docker demo with simulated data at localhost:3000
- Gives: Tier 1: a source build with tests and a proof check
- Gives: Tier 2: hand-off to flashing and setting up a live board
- Gives: Warnings about unsupported boards and accuracy limits
- Gives: A table of next steps
- Say: "/ruview-start docker"

### ruview-hardware-setup
- Asks for: Board type and flash size (8 MB or 4 MB)
- Asks for: The serial port the board is on
- Asks for: Your WiFi name and password and the address of the computer that collects data
- Gives: Firmware built and flashed
- Gives: WiFi and target address written to the board
- Gives: A 15-second serial readout to confirm data
- Gives: A table of common errors and fixes
- Say: "/ruview-flash 4mb COM8"

### ruview-configure
- Asks for: Serial port, WiFi name and password, target address and node number
- Asks for: Optional: channel, transmitter filter, several-board slots, detection thresholds
- Gives: Board settings written over the serial port (or a dry-run file)
- Gives: An explanation of the trade-offs of each option
- Gives: A check that data arrives at the collector
- Say: "/ruview-provision --port COM8 --ssid MyWiFi --password ... --target-ip 192.168.1.20"

### ruview-applications
- Asks for: Which app you want, or your situation (through a wall, a person lying still, search and rescue)
- Gives: The command that starts the app (live board or Docker demo)
- Gives: Advice on which sensing mode fits the situation
- Say: "/ruview-app vitals"

### ruview-model-training
- Asks for: Which track: no-camera pose, camera-supervised pose, embeddings, cross-room transfer, on-device adaptation or cloud GPU
- Asks for: For camera-supervised pose: a webcam and paired recordings
- Gives: A trained model file (model.rvf)
- Gives: An accuracy score for pose (PCK@20)
- Gives: The test and proof checks after training
- Gives: Optional: a cloud GPU run and a Hugging Face upload
- Say: "/ruview-train camera-free"

### ruview-advanced-sensing
- Asks for: Which feature you want to use
- Asks for: Two or more boards for the multi-viewpoint features
- Gives: The source modules and design notes for that feature
- Gives: The tests to run after a change
- Say: "/ruview-advanced cross-viewpoint"

### ruview-verify
- Asks for: Which checks: tests, proof, bundle or all (default all)
- Gives: Test results
- Gives: A VERDICT PASS or FAIL from the proof check
- Gives: A witness bundle archive in dist/ with a self-check script
- Gives: The 12-point pre-merge checklist
- Say: "/ruview-verify all"

## Know before installing
- A tool server starts by itself: when the plugin is active, Claude Code downloads and runs the rvAgent package from npm, with no version pinned.
- Re-provisioning a board erases every setting you do not pass again.
- Only some boards work: the original ESP32 and the ESP32-C3 are not supported.
- Small clone boards can overheat: the skill tells the assistant to ask about the board before flashing.
- Windows-first paths: the build steps use the author's Windows paths and COM8 as the default port; confirm your own.
- Not a medical device: vital signs and pose need your own checks before real use.

## Install (official commands)
- Claude Code: `/plugin marketplace add ruvnet/RuView` && `/plugin install ruview@ruview` — Send as two separate prompts. Try without installing: claude --plugin-dir ./plugins/ruview from a clone.
- Claude Code (live pane mod, early access): `/plugin install ruview-live@ruview` — Optional. /ruview opens a read-only pane with your boards and radar. If /ruview is missing, start Claude Code with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1.
- Codex: `mkdir -p ~/.codex/prompts` && `cp plugins/ruview/codex/prompts/*.md ~/.codex/prompts/` — Run from a clone of the repo. Optionally link plugins/ruview/codex/AGENTS.md as the project AGENTS.md. Codex gets the commands only, not the skills or agents. Remove: `delete the ruview-*.md files from ~/.codex/prompts/`.

## What runs on the user's machine
- Code on the machine: yes (Claude Code plugin installs (the Codex prompts are text))
- Network: True
- Plugin activation in Claude Code: Starts an MCP server with npx -y @ruvnet/rvagent (downloaded from npm, version not pinned). It reads from a sensing server at http://localhost:3000 by default; RVAGENT_SENSING_URL changes it.
- When you run a command or skill: The skills tell your assistant to run the project's own tools: docker pull/run, cargo builds and tests, Python and Node scripts in the RuView checkout, flashing over the serial port. Skills are allowed Bash, Read, Write and Edit.
- /ruview-train gpu or publishing: Tells your assistant to log in to Google Cloud and run training on a rented GPU machine, or to upload a model to Hugging Face.
- /ruview-verify: Suggests npx @claude-flow/cli@latest security scan after security changes, and writes a witness bundle to dist/.
- ruview-live mod, while its pane is open: A Claude Code mod runs the @ruvnet/ruview command-line tool on a timer: it listens briefly on a UDP port for board data and, if you set a radar host, reads an ESPHome radar on your network. Read-only.
- Install: plugins/ruview/scripts/smoke.sh is a structure check of the plugin files for maintainers; it is not run on install.

## Settings
- `RVAGENT_SENSING_URL`: Sensing server the rvAgent tools read from (default http://localhost:3000)
- `CLAUDE_CODE_ENABLE_FUNCTION_HOOKS`: Set to 1 if /ruview from the ruview-live mod is missing
- `ruview-live userConfig`: udpPort (5005), radarHost (empty skips radar), refreshSeconds (15), captureSeconds (3), liveRefreshSeconds (4)

## Examples (from the pack's own files)
- The /ruview-verify witness bundle contents: output format defined in plugins/ruview/skills/ruview-verify/SKILL.md

## Author's numbers (as reported, not run by AgentGid)
- Camera-supervised pose training reaches 92.9% PCK@20 in about 19 minutes on a laptop (Webcam + MediaPipe as ground truth paired with ESP32 data (ADR-079)). Source: ruvnet/RuView/plugins/ruview/skills/ruview-model-training/SKILL.md
- Camera-free pose training takes about 84 seconds on an M4 Pro, with modest accuracy. Source: ruvnet/RuView/plugins/ruview/skills/ruview-model-training/SKILL.md
- The Rust workspace has 1,400+ passing tests. Source: ruvnet/RuView/plugins/ruview/skills/ruview-verify/SKILL.md

## Steps only the person can do
- Connect an ESP32-S3 or ESP32-C6 board by USB and tell the assistant its serial port (plugins/ruview/skills/ruview-hardware-setup/SKILL.md) Then: The assistant builds, flashes and provisions the board
- Give your WiFi name and password for provisioning (they are written to the board) (plugins/ruview/commands/ruview-provision.md) Then: The board joins your WiFi and streams to your computer
- Log in to Google Cloud with your own account and project (only for cloud GPU training) (https://cloud.google.com) Then: The assistant runs scripts/gcloud-train.sh; the machine is billed to your account
- Log in to Hugging Face (only to publish a model) (https://huggingface.co) Then: The assistant runs the publish script

## Mistakes to avoid
- Not: Re-provision a working board with only the setting you want to change. Instead: Pass the full set of settings every time, or use --force-partial knowingly; read the current values from the boot log first. Why: Provisioning replaces the whole settings area on the board
- Not: Buy an original ESP32 or ESP32-C3 for CSI sensing. Instead: Use an ESP32-S3 (8 MB or 4 MB) or an ESP32-C6. Why: Single-core chips cannot run the signal pipeline
- Not: Build the firmware in Git Bash on Windows. Instead: Use the Python subprocess pattern from the skill with MSYSTEM variables removed. Why: ESP-IDF v5.4 does not support MSYS2 or Git Bash
- Not: Test firmware changes in mock mode. Instead: Test with real WiFi data on a real board. Why: A threshold bug only showed up with real data
- Not: Install the plugin into an unrelated project and expect the commands to work. Instead: Use it inside a clone of the RuView repository. Why: The commands assume RuView's folders (v2/crates, firmware/esp32-csi-node, scripts)

## What leaves the machine
- A download request for the @ruvnet/rvagent package → The npm registry, Each time the plugin's MCP server starts in Claude Code (npx -y, no pinned version). Opt out: Disable the rvagent MCP server, or use the Codex prompts, which start nothing
- The Docker image ruvnet/wifi-densepose:latest → Docker Hub (download), When you choose the Docker demo. Opt out: Build from source instead
- Training data and job, run on a rented GPU machine → Google Cloud, Only with /ruview-train gpu. Opt out: Train locally (the default tracks run on a laptop)
- A trained model file and its card → Hugging Face, Only when you ask to publish. Opt out: Do not run the publish step

## Lines in the pack aimed at agents (ask the person first)
- "npx @claude-flow/cli@latest security scan            # after security-related changes" (plugins/ruview/skills/ruview-verify/SKILL.md): Unpinned @latest package from npm. Ask the person before running it.
- "Project `cognitum-20260110` has L4 / A100 / H100 quota." (plugins/ruview/skills/ruview-model-training/SKILL.md): This is the author's own Google Cloud project. Never set it; ask the person for their own project and confirm the cost first.
- "env['IDF_PATH'] = r'C:\Users\ruv\esp\v5.4\esp-idf'" (plugins/ruview/skills/ruview-hardware-setup/SKILL.md): The build command has the author's machine paths. Replace them with the person's own ESP-IDF paths.
- "python archive/v1/data/proof/verify.py --generate-hash" (plugins/ruview/skills/ruview-verify/SKILL.md): This rewrites the expected proof hash so the check passes again. Only after the person agrees the change is legitimate.

## May I use it like this?
- commercial-app: yes (MIT: keep the license notice)
- redistribute-files: yes (MIT: keep the copyright and permission notice in copies)
- client-work: yes (MIT)

## Which one to pick
- If You work in the RuView repository and want guided setup, sensing apps, training and verification: skill:ruvnet/RuView
- If You only need to flash or talk to any Espressif chip, outside RuView: tool:esptool
- If You want pose detection from a camera rather than from WiFi signals: tool:mediapipe

## Works well with
- agent:claude-code: The skills, commands, agents and the rvAgent MCP server are a Claude Code plugin
- agent:openai-codex: The seven commands are mirrored as Codex prompts with an AGENTS.md
- tool:esptool: The hardware skill flashes the ESP32 with esptool as an alternative to idf.py
- tool:mediapipe: Camera-supervised pose training uses MediaPipe as ground truth

## Works in and needs
- Works in Claude Code: plugin
- Works in Codex: prompt files + AGENTS.md
- Needs A clone of the RuView repository: Every command runs inside its folder layout
- Needs Node.js (npx): The rvAgent MCP server starts with npx
- Needs Docker, Rust (cargo), Python: Demo, builds, tests and scripts
- Needs ESP-IDF v5.4 and an ESP32-S3 or ESP32-C6 board: Only for live sensing
- Needs Google Cloud account: Only for cloud GPU training

## How to tell it works
Install the plugin in Claude Code and type /ruview-start docker. Your assistant should give the Docker command for the simulated demo and point you to http://localhost:3000, then list next steps such as /ruview-app and /ruview-flash.

## Related
- ruvnet/rvcsi

## Safety
The plugin starts an npm package (@ruvnet/rvagent) without a pinned version, and /ruview-verify suggests npx @claude-flow/cli@latest. The skills let the assistant run shell commands, flash boards and, if you ask, rent cloud GPUs. Commands contain the author's Windows paths and port; read them before approving.
