# Tensor Dynamics Watch assistant pilot

This pilot helps you use Watch and checks a running **Watch 0.39.0** on this
computer. It does not yet install Watch, add a camera, create an API key, change
settings, or deploy an integration. It is not a public ChatGPT directory listing.

## ChatGPT desktop and Codex on macOS, Windows and Linux

1. Install and sign in to the supported desktop AI app. This pilot also needs
   **Node 20 or newer** and the **Codex CLI**. Use the official Node website and
   [Codex installation guide](https://developers.openai.com/codex/cli) if missing.
2. Extract the downloaded ZIP, then choose the launcher for your operating system:
   - macOS: open **Install Watch Support.command**.
   - Windows: open PowerShell in the extracted folder and run `node .\scripts\install-local.mjs`. The included **Install Watch Support.ps1** runs the same command when your normal script policy permits it. No policy bypass is required.
   - Linux: open a terminal in the extracted folder and run `sh "Install Watch Support.sh"`, or `node scripts/install-local.mjs`.
   Keep the window open if it reports a missing prerequisite.
3. Start a new local ChatGPT or Codex chat. Find **Tensor Dynamics Watch Pilot**
   in Plugins and select **Watch Support Pilot**. Restart the desktop app if it
   does not appear. On Linux, use Codex CLI; a native ChatGPT desktop app is not required or claimed. Ordinary web chat does not run this local connector.
4. Open Watch 0.39.0. The connector uses Watch's existing local pairing
   automatically. If it reports missing or disabled access, contact pilot support
   to repair the local pairing; some builds have no visible MCP settings page.
   Your credentials stay on this computer; do not paste them into chat.
5. Say: **“Set up Watch with me. Check what is installed and guide me through
   the next step.”**

If macOS blocks the Watch pilot or installer, stop and contact support. Do not
disable operating-system protection to run a pilot. The current Watch Mac pilot
still needs signing/notarization work before broad customer distribution.

## Claude Desktop on macOS and Windows

Download the separate **Claude Desktop MCPB** from the
[Watch assistant page](https://watch.tensordynamics.org/docs/setup-assistant#desktop-pilot).
Open that file in Claude Desktop and review the requested access. Claude supplies
its own Node runtime. Keep the default local address unless Watch shows a different
port, then ask it to check Watch. The MCPB is a separate download, not inside this ZIP. No official Claude Desktop Linux distribution is claimed; use the portable connector with Codex CLI on Linux.

## Platform validation

The portable connector and launchers are provided for macOS, Windows and Linux.
Mac clean-profile installation and real Watch connectivity are tested. Windows
npm-shim resolution and paths with spaces have isolated tests; its PowerShell
launcher and real Windows Watch setup still require a Windows machine. The POSIX
launcher is syntax-checked; a complete Linux Watch/client installation still
requires Linux validation. These checks are not a full OS installation pass.
Automatic Watch installation/configuration is not included in this public pilot;
the action-enabled FDE development work has separate acceptance results.

## What to expect

The assistant can check the app version and authenticated connection, read current
guides, list active camera status, and read requested recent event summaries.
Its guides cover installation, activation, cameras, settings, missions, 360 viewing,
supported schedules, event/database review, notifications, and public read-only
Event API integration. It walks you through the documented controls in Watch.
It can also inventory the local system, calculate theoretical TOPS, observe the
current workload's performance, and prepare a capacity test plan.
Requested summaries are shared with your AI assistant. Images, camera addresses,
and credentials are excluded from results. An empty active-camera list does not
mean there are no saved cameras.

A healthy app does not prove that camera video, analysis or notification delivery
works. Follow the assistant's checks in Watch. Enter license keys, API keys and
camera passwords directly in the appropriate local app. This pilot never creates
or changes those credentials.

For hardware and capacity, say: **“Assess this computer for 8 cameras, one analysis
per camera every 5 seconds, and results within 10 seconds.”** The assistant will
ask for the actual video settings and workload when needed. With Watch running,
ask it to observe the current cameras for 30–45 seconds. Too few or stale metrics
leave capacity unknown. TOPS is theoretical; short observations are provisional
and do not certify long-term camera capacity or physical camera-to-alert latency.
Hardware strings and performance numbers go to your AI conversation; serials,
user/host identity, images and internal analysis-service choices are excluded.

For a technical installation from this extracted folder, run
`node scripts/install-local.mjs`. Re-running preserves the previous connection
address. Use `--watch-url http://127.0.0.1:PORT` only to explicitly change the port.

Support: https://watch.tensordynamics.org/contact