Tools · Meta Ray-Ban Display
Bring your own computer and coding agents to your glasses, with private connections and a complete Windows setup guide.
Also known as Agent Hub Glasses
A closer look
Agent Hub connects your display to coding agents on a Windows computer you control. Choose a computer, provider, project and conversation, then send prompts and follow the work from your glasses. The public app starts with no configured computers.
This is a bring-your-own-computer setup: you install the clean starter, sign in to your own providers and create a private connection. No creator computer, account, project, conversation, GitHub service or provider credit is included. Start with one provider; you can add more later.
Make it your own
ZIP for your own computer. Includes the server, setup guide and a blank configuration example. Install instructions are included inside.
Install Node.js LTS 22 or newer using the official link below. Open Windows Terminal with PowerShell and check the versions below. Use the same Windows account for the starter and your agent logins. This walkthrough uses native Windows commands; do not mix a provider installed only inside WSL with this Windows hub. Git for Windows is recommended for Git projects and Claude Code’s Bash tools.
node --version
npm.cmd --versionDownload the starter ZIP above, extract all files into a folder you own, then open PowerShell in that folder. Run setup once. It creates .env with a fresh 256-bit BRIDGE_TOKEN and will not overwrite an existing file. The hub itself needs no npm install. Open .env locally when configuring providers; this token is the hub password, not a provider API key. Keep the file private.
npm.cmd run setup
notepad .envFollow the matching provider section below and complete its login on this computer. For Codex or Claude Code, install the pinned local adapter below and set BRIDGE_AUTOSTART=1 in .env; this starter targets adapter version 0.7.7. Its package name refers to a shared local adapter, not a different glasses app. For OpenCode or local models, leave BRIDGE_AUTOSTART=0 unless you also want Codex or Claude Code. Keep PROXY_HOST=127.0.0.1, HUB_AUTOSTART=0 and HUB_DISTRIBUTION=starter.
npm.cmd install -g @evenrealities/even-terminal@0.7.7Return to the extracted starter folder and start the hub. Leave this terminal open and keep the PC awake. The hub listens locally on port 3457; provider adapters use separate local ports. An installed provider is not the same as a running, signed-in provider. The next steps make this authenticated hub reachable from your glasses.
npm.cmd startInstall Tailscale on this computer and sign in to your own account. In a second PowerShell window, run the Funnel command below and approve any Tailscale setup prompt. Copy the HTTPS address it reports. Funnel makes this port reachable over the internet, while Agent Hub still requires your token. Use Funnel rather than a tailnet-only Serve address. Route only port 3457; do not expose the local agent-adapter ports.
tailscale funnel --bg 3457In PowerShell, return to the starter folder and run these four commands in order; stop if any command reports an error. Paste your own Funnel HTTPS address when asked and give this computer a short name. The commands validate the address and the token generated by setup, then copy your private Agent Hub link to the clipboard without printing it. This link is a password: never post it publicly. No details are sent to GlintShelf.
$hub = Read-Host 'Paste your HTTPS hub URL'; $u = [uri]$hub; if (!$u.IsAbsoluteUri -or $u.Scheme -ne 'https' -or $u.UserInfo -or $u.Query -or $u.Fragment) { throw 'Use a plain HTTPS hub URL' }
$token = (Get-Content -LiteralPath .env | Where-Object { $_ -match '^BRIDGE_TOKEN=' }) -replace '^BRIDGE_TOKEN=', ''; if ($token -isnot [string] -or $token -notmatch '^[0-9a-f]{64}$') { throw 'Run setup to create a valid local token' }
$n=Read-Host 'Computer name'; if(!$n.Trim() -or $n.Length -gt 80 -or $n -match '[\x00-\x1f\x7f]'){throw 'Invalid name'}; $bundle=@{version=1;hosts=@(@{name=$n;baseUrl=$u.AbsoluteUri;token=$token})}|ConvertTo-Json -Depth 4 -Compress
Set-Clipboard ('https://claude-code-glasses.onrender.com/#connections=' + [uri]::EscapeDataString($bundle))Transfer the copied private link to your paired phone using a private method you trust. In Meta AI, open App Settings → App Connections → Web Apps → Add a Web App. Name it Agent Hub and paste the full private link as its app address, then Connect. Open that Agent Hub entry on the glasses once; it imports the connection and shows Computers. The public QR only opens blank setup. Connecting in a phone browser alone does not configure the glasses. Menu wording can vary with Meta AI version.
On the computer, start your chosen agent in an existing project folder and finish its login and initial setup. On the glasses choose Hub → Computers → your computer → agent → project → session. Refresh if needed, or choose New session after selecting the project. Send a small prompt and verify the reply before granting broader tasks. Keep ASK ME approvals enabled. Updates to the standalone starter are manual: preserve your private .env, stop the old hub with Ctrl+C, and start the reviewed replacement.
Configure only the providers you want to use on your own computer.
Run the install command, then open PowerShell in the existing project folder you want to use and run codex.cmd. Complete OpenAI’s sign-in flow with your own account and create a session in that folder. A desktop-only or WSL-only installation may not put the CLI on this Windows PATH; codex.cmd --version must work here. Enable the pinned Claude/Codex adapter in step 3. Provider access and usage limits belong to your account.
npm.cmd install -g @openai/codex
codex.cmd --version
codex.cmdInstall Claude Code using the official guide or the WinGet command below, then reopen PowerShell so PATH changes apply. In your own project folder, run claude and finish its first-run login. Git for Windows is recommended for Bash-based tools. Enable the pinned Claude/Codex adapter in step 3. Use your own Claude Code account or supported provider setup; access and billing are separate from Agent Hub.
winget install Anthropic.ClaudeCode
claude --version
claudeInstall the native Windows CLI and sign in to a model provider with your own credentials. Run opencode.cmd inside your project folder once to establish a session. In another terminal, start its local server with the final command below and leave it running. Keep OPENCODE_AUTOSTART=0 for this manual setup. If you already protect OpenCode with a server password, supply the matching OPENCODE_SERVER_PASSWORD to the hub’s private .env as well. Only the authenticated hub port is exposed through Funnel.
npm.cmd install -g opencode-ai
opencode.cmd auth login
opencode.cmd
opencode.cmd serve --hostname 127.0.0.1 --port 4096The starter also includes adapters for Ollama, Unsloth Studio, OpenClaw and Hermes; follow the starter README and each provider’s own installation guide. Ollama must be running locally with a model available. Unsloth Studio needs a loaded model and your own API key. OpenClaw needs its Gateway, and Hermes needs its CLI on PATH. These packages, models and accounts are not bundled. Begin with one working provider before adding another.
Keep the PC awake, the npm.cmd start terminal running, and Tailscale signed in. Run tailscale funnel status on that PC to check the HTTPS route points to port 3457. Confirm that you used the public Funnel address, not localhost, a private IP or a Serve-only address. After enabling Funnel, DNS and certificates can take time. Refresh Computers after the route is ready.
Use the token from this exact starter folder’s .env. It must match the hub currently running there. The link generator expects the 64-character token created by npm run setup. A provider API key is not a hub token. After rotating a hub token, restart the hub and generate a new private connection link for your devices. Copy the entire link, including its #connections fragment; a rejected import preserves existing saved connections.
Read the explanation on the Agents screen. Make sure the provider CLI works in native Windows under the same user as the hub, and that its own login succeeds. Codex and Claude Code need the pinned even-terminal adapter plus BRIDGE_AUTOSTART=1. OpenCode needs its server on local port 4096. Restart the hub after editing .env. If PowerShell blocks npm.ps1, use the documented npm.cmd commands instead of weakening system execution policy.
Start a real session in the intended folder using that provider on the computer, then refresh the glasses’ project and session lists. Each computer and provider has its own list. Keep the project folder available at its original path. If a Codex session is still active in another app, finish or release that session there before resuming it from the glasses; do not open competing writers to the same session.
Repeat the PC setup with a separate token and HTTPS address for each computer. The private link above imports one connection; the app can merge imports and preserve other saved connections in the same browser. For a fresh installation containing several computers, an advanced #connections bundle can contain up to eight host records. If Meta AI allows editing the existing app’s address, use that entry for a later import and open it once on the glasses. Editing availability varies; phone browser imports alone never update glasses storage.
The clean starter connects agents running on your own PC. Those agents can work in your own locally checked-out GitHub projects using your configured permissions. An always-on GitHub Direct hub is a separate advanced backend, not included or provisioned by this download. It requires your own hosting, repository authorization and provider configuration. The public app supplies no creator GitHub connection; do not substitute a GitHub token for the PC hub’s BRIDGE_TOKEN.
Built for your own setup
Screenshots show the prepared Agent Hub interface and clearly labeled fictional provider examples. No real account, host or token is shown.