Three steps: run the server on your Mac, connect the app on your iPhone, say your first prompt. No decisions required up front — everything is tunable later.
Check these before you start
Even G2 glasses paired with the Even Realities app on your iPhone
A Mac that stays on with Node.js 20.11+ — run node -v first; newer compatible major versions are fine, so do not install node@20 beside one
Claude Code and/or Codex CLI on that Mac — Claude Desktop is not the terminal CLI. Install Claude Code on one line with npm install -g @anthropic-ai/claude-code (never use sudo), then run claude and finish sign-in. Claude Code setup powers Opus, Fable, and Sonnet; Codex CLI powers GPT Frontier and GPT Balanced. Install both if you can: the server detects and uses whichever you have, and you can switch per query.
whisper.cpp and about 6 GB free for voice prompts — run brew install whisper-cpp once. The setup command verifies disk space before downloading anything. This is required for local voice; text chat works without it.
Local spoken replies on Apple silicon — optional: run brew install python@3.12 ffmpeg espeak-ng. Python 3.11 or 3.12 is supported. The normal server start provisions Kokoro asynchronously; chat and transcription work while it prepares.
ffmpeg for photos — the command above also covers phone photos, answer-image galleries, and manual 288×144 G2 image viewing. Text and transcription still work without it.
Came here from the Even Hub? You already have the app half — the server in Step 1 below is the missing piece.
Easiest path: COS Control. A menu bar app for the Mac that installs the server for you, keeps it running, launches at login, and updates and rolls back safely. No terminal to babysit. Step 1 below starts with it; the terminal command stays as the alternative. Download COS Control (macOS 14 or later, Apple Silicon).
Three steps, smart defaults
Get the glasses live now; tune anything later in Settings. This setup is the same whether or not you already run a COS.
ModelSonnet (swap anytime)
VoiceBalanced local (Small preview · Turbo commit · Large-v3 HQ)
On the lensMinimal (clean default)
ContextInherits your COS (if you run one)
1 · On your Mac — run the server
The easy way: COS Control. A menu bar app that installs the server, keeps it running, launches at login, and updates and rolls back safely. Most people should start here.
Download COS Control (macOS 14 or later, Apple Silicon), unzip it, and move it into your Applications folder. The Control page has the checksum and the details.
On the first open macOS says it could not verify the developer. That is expected: COS Control is open source and not notarized yet. Open System Settings, Privacy & Security, and click Open Anyway.
Run Guided Setup, choose Balanced, and click Install Server when it is offered. Control checks whisper-cpp and free disk space, provisions the local voice models, then starts the server as a background service.
Once the menu bar shows the server running, click Copy Pairing Token. That is the API token for Step 2 (it stays on the clipboard for 60 seconds).
Or the terminal. The same server, running in the foreground. Keep the window open while you use the glasses.
$ npx --yes @gotcos/glasses-server@latest --setup-transcription --transcription-tier balancedclick to copy
The terminal prints your server URL (Tailscale and Wi-Fi addresses, labeled) and your API token — you'll paste those into the app in Step 2. The token is saved to ~/.cos-glasses/.env. For transcription, install whisper-cpp first and have about 6 GB free; setup verifies space, then provisions Small.en preview, Turbo commit, and Large-v3 HQ models. For optional local spoken replies on Apple silicon, run brew install python@3.12 ffmpeg espeak-ng; the server validates and self-repairs its pinned Kokoro environment.
2 · On your iPhone — install and connect
Open the Even Realities app → Even Hub → install COS Glasses.
Open it and enter the server URL and API token (or tap "I have a COS username" and skip the typing). The URL is http:// your Mac’s address :3141: the terminal prints it; with COS Control, use the Mac’s Wi-Fi address from System Settings or its Tailscale address. The token is what the terminal printed, or what Copy Pairing Token put on the clipboard.
Tap Connect.
3 · Prove it
If the terminal showed LOCAL VOICE NOT READY, or Guided Setup said whisper-cpp is missing, run brew install whisper-cpp, then rerun the server command (Ctrl-C first) or run Guided Setup again in COS Control before tapping the mic.
Tap the mic and ask anything — try "What should I focus on today?"
Watch the answer stream onto the lens in green. That's your COS, live on your face.
Learn tap-and-hold: tap and release, then quickly press again and keep holding for the shortcut window above the HUD. Scroll to Ask COS, Model, Start Meeting, Messages, Sessions, Tasks, or Home, then tap to select. While recording, Resume Meeting and Stop Meeting lead; Tasks stays in Quick Actions instead. Practice the hold gesture.
Practice the page-specific controls: in a reader, tap once for footer choices, scroll, then tap again to confirm. This is not the quick double-tap that starts a referenced reply. Tasks and Sessions have rolling footer menus; task wrapping needs glasses 6.9.455+. Try Messages, Sessions, Tasks, and Ask COS with the simulated ring.
Optional photo proof: tap the camera, stage one photo, ask "What am I looking at?", then open its phone gallery. Turn on Lens images to test the manual G2 preview.
Already run a COS from the Starter Kit? Run the command in Step 1 from your COS folder — the server detects it and glasses chat loads that brain automatically. New to COS? The glasses run great on their own; set up the full system anytime at gotcos.com.
Already have meeting files? After setup, open COS Control and choose Meetings Library. The recommended path is simply the folder that already contains your YYYY-MM/*.md meeting folders. If you keep several libraries, choose their parent; each custom-named folder only needs meetings/YYYY-MM/*.md inside. The website does not change your Mac; the final folder choice happens in COS Control. You can also skip this step and use new standalone G2 recordings.
Never babysit the server again. Have your COS create a /cos-glasses skill — ask it: "add the /cos-glasses skill from the Starter Kit." Then /cos-glasses starts the server, health-checks it, keeps it on the latest release, and troubleshoots connection issues for you. Grab it at skills/cos-glasses.md.
Pick up where the glasses left off. Add the /glasses-day skill and your COS pulls the day's glasses questions and answers into your desk session — so anything you asked on the go becomes something you can act on. Grab it at skills/glasses-day.md.
Nothing here blocks setup — the glasses run great on the defaults. Tune when you're ready; every pick renders a personalized checklist at the bottom.
How will you use them?
Pick all that apply. Everything below tunes to your answer.
Pick your default brain
A soft default, not a commitment — every model stays available per query, by voice or from the toolbar. Opus, Fable, and Sonnet follow Claude Code; GPT Frontier and GPT Balanced follow the newest capable Codex generation automatically.
No per-token cost on chat. Chat runs on the Claude Code and/or Codex plans you already have. Install both engines and all five brains stay available — the server uses whichever is present. Voice is separate; see Compute below.
How much can it touch? On the Claude brains COS runs with full tool access by default. It can read and write files and run commands on the server Mac without prompting you. That is what makes the glasses useful for real work. Prefer a tighter posture? Set COS_CLAUDE_TRUST_MODE=allowlist in the server environment before you launch. The query keeps web search and read-only access to your workspace, while shell, edits, and writes are denied. Needs server 6.41.0 or newer. Earlier servers denied workspace reads too, so a hardened install answered "I don't have access to your workspace files." Either way COS instructs the model to treat web pages, files, and transcripts as data, never as instructions.
Want a local model too? Install Ollama on the server Mac and pull a model. COS detects it and offers it automatically as a local option beside the brains above, and the model runs entirely on your Mac. Optional: with no Ollama installed there is nothing to configure.
Who is it working for?
Set the name of the person COS serves. It personalizes answers and the Knowledge graph’s starting point. You can change it in COS Control → Memories → Memory settings → Owner name. This is separate from the Mac that indexes your documents.
This directly improves transcription. COS feeds these spellings to the decoder. Factory examples are ignored, so an untouched profile can no longer bias Whisper toward literal placeholders such as NameOne or YourCompany.
What context should it see?
Wire the sources you live in. We pre-select what fits your use cases. Adjust freely.
Your role is not your folder structure. Executive, marketer, engineer, product lead, people lead, attorney, chief of staff, clinician, or something else: the Blueprint personalizes how COS works, never what your folders must be called. For the easiest setup, choose the folder that directly contains YYYY-MM/*.md. Organize several libraries only if you already work that way; their names are entirely yours.
What should it remember?
Two stores, both plain markdown on your own machine, both readable from the lens.
Memory
Individual notes worth keeping: a decision and why, a correction you had to make twice,
a fact about a person. One file per note. On the glasses you browse them by type and
summary, and double-tap one to ask a follow-up grounded in that exact record.
memory/2026-08-09-pricing-call.md
Threads
Work that spans weeks: a launch, a hire, a renewal. One file per thread, appended as it
moves, so the whole arc stays in one place instead of scattered across meetings.
threads/website-rebuild.md
No database required. Since server 6.22.0 a folder of markdown
is a complete setup. Any nesting, any filename, front matter optional. A symlink is followed, so
notes that already live somewhere else can stay there — at any depth on
6.22.1, top level only on 6.22.0. Add a Python bridge later and the same
files become what it indexes — nothing to migrate, because the files were always the source.
How do you want to start?
Where the notes go. Pick the folder in COS Control and it shows you the resolved path, so you always know where a note you wrote will land.
Where does it run, and how does voice work?
The server runs on your Mac. Your keys, your machine. Pick how voice is transcribed.
Live transcription tier
Spoken replies are separate from transcription. The server provides Local (Kokoro), Local-first, and OpenAI speech engines. Explicit Local fails closed and never calls OpenAI. Local-first can fall back only after a local failure and only when your own OpenAI key and budget are configured. Check tts_local.ready in /api/health.
Balanced is the recommended setup. Small.en produces provisional prompt words quickly, Large-v3-Turbo commits the authoritative live transcript, and Large-v3 polishes saved work. Max is opt-in and reuses the existing Large-v3 worker for live commit. A missing Max model falls back visibly to Turbo instead of breaking transcription.
Phone to Mac: connect both over a secure mesh (Tailscale recommended) so the glasses reach your server from anywhere. The setup below covers it.
Photos stay local: install ffmpeg on the Mac for normalized phone uploads, private answer-image galleries, and exact 288×144 G2 previews. Images never need a COS-hosted cloud.
What shows on the lens?
Your default layout. Adaptive switches it by context.
Lens images are manual and off by default. Turn them on from the phone companion. Photo messages end with a cover page; a deliberate two-tap confirmation opens the image. Meetings and microphone-critical states always win.
Your COS Glasses, ready to set up
Make your picks above and your personalized setup appears here.
ModelSonnet (default)
VoiceLocal whisper (free)
Context wired·
Memory & ThreadsNew folders
On the lensMinimal
1 · Start your server (self-host)
$ npx --yes @gotcos/glasses-server@latest --setup-transcription --transcription-tier balancedclick to copy
The app half installs from the Even Hub on your iPhone — covered in Step 2 of the setup above.
Even G2 smart glasses paired with the Even Realities app
Claude Code and/or Codex CLI on your Mac — install both if you can; the server uses whichever is present. Claude Desktop is not the CLI. For Opus/Fable/Sonnet, run npm install -g @anthropic-ai/claude-code without sudo, then run claude and finish sign-in. GPT Frontier/Balanced need Codex.
COS Control (recommended) — the Mac menu bar app that installs and runs the server for you, launches at login, and updates and rolls back safely. macOS 14 or later, Apple Silicon. Download
ffmpeg for phone and answer images — brew install ffmpeg
Tailscale (recommended) — a secure mesh so your phone reaches the server on your Mac. Download
Why Tailscale? Your glasses app on the phone needs to reach the server on your Mac. Tailscale gives both a stable IP that works from anywhere — no port forwarding, no firewall hassle.
Start the COS Glasses server
Runs on your own Mac, as a background service under COS Control or in a terminal you keep open.
No terminal: COS Control.Download COS Control, move it into Applications, approve the first open under Privacy & Security, then run Guided Setup and click Install Server when it is offered. It installs this same server, starts it, and keeps it running at login; Doctor, safe restart, update and rollback live in the menu bar. Skip the command below, and use Copy Pairing Token in step 4. Already running the server in a terminal? Stop it with Ctrl-C first; Control refuses to take over a process it does not own.
Or in the terminal:
$ npx --yes @gotcos/glasses-server@latestclick to copy
Voice runs locally. Mac whisper.cpp remains the authoritative meeting engine and costs zero per minute. Install once: brew install whisper-cpp. Since COS Glasses 6.8.336, you can also choose on-phone Moonshine Tiny or Base for rolling Message and Meeting preview. Text forms and may revise while you speak; the shared glossary + AI finalizer cleans settled phrases, while speaker-attributed Large-v3 remains canonical for meetings. This requires server 6.27.2 or newer. Cloud fallback stays off unless the server operator explicitly enables it with COS_OPENAI_WHISPER_FALLBACK=1 and configures a key.
Running the full COS? Add the /cos-glasses skill so your agent starts the server, health-checks it, keeps it on the latest release, and troubleshoots for you — ask your COS to "add the /cos-glasses skill from the Starter Kit" (skills/cos-glasses.md).
Connect your devices
Install Tailscale on your Mac and iPhone; sign in with the same account on both.
Open Tailscale on your Mac and note its IP (starts with 100.).
Same WiFi? You can use your Mac's local IP instead — Tailscale is just more reliable across networks.
Install COS Glasses on the Even Hub
Open the Even Realities app on your iPhone.
Go to Even Hub, search "COS Glasses", install, and open it.
Connect and go
Enter your Mac's Tailscale IP and paste the API token: the server terminal prints it, or click Copy Pairing Token in COS Control (60 seconds on the clipboard).
Leave the server running, in the terminal or under COS Control, then tap Connect.
Have a COS username? Tap "I have a COS username" — your server details resolve automatically, no IP or token needed.
Switch models anytime: open Settings, choose Default chat model, tap the toolbar model button, or say "switch to GPT Frontier."
Troubleshooting
"Cannot reach server" — confirm both devices are on Tailscale (or the same WiFi) and the server is still running (the terminal window, or the COS Control menu bar item).
"Claude not found" — Claude Desktop is not the terminal CLI. Run npm install -g @anthropic-ai/claude-code on one line without sudo, then run claude and finish sign-in. Verify with claude --version.
npm EACCES / root-owned cache — do not rerun with sudo or recursively change system ownership. Use npm_config_cache="$HOME/.cos-glasses/npm-cache" npx --yes @gotcos/glasses-server@latest. Server 6.12.2+ does not run a second install inside npm's temporary cache.
@latest launched an older version — compare /api/health with npm view @gotcos/glasses-server dist-tags.latest, then restart with npm_config_prefer_online=true npm_config_cache="$HOME/.cos-glasses/npm-cache" npx --yes @gotcos/glasses-server@latest.
"Codex not available" — GPT Frontier/Balanced need the local Codex CLI on the server Mac; run codex --version, then codex login.
Transcription unavailable — run brew install whisper-cpp for free local transcription. Cloud fallback remains off unless an administrator sets both COS_OPENAI_WHISPER_FALLBACK=1 and OPENAI_API_KEY.
Local spoken replies unavailable — on Apple silicon run brew install python@3.12 ffmpeg espeak-ng, restart COS, and allow first-run Kokoro provisioning to finish. Confirm tts_local.ready in /api/health. Explicit Local fails closed; choose OpenAI only if you intend to use your key.
Photos unavailable — run brew install ffmpeg, restart with npx --yes @gotcos/glasses-server@latest, and confirm Lens images is on before testing outside a meeting.
Server disconnects — it auto-reconnects; if persistent, restart with npx --yes @gotcos/glasses-server@latest.
What it feels like
COS Glasses · Personalize
Build your COS Glasses.
Your Even G2, running as a chief of staff on the lens. The context you have already built, now it follows you off the screen.
No slide deck. No black box. You own it, you do not rent it.
REC0:12 ★1 100%
[Priya] if we shift budget to the comparison pages we can hold CPA under target through July.
[You] Lock that in. Flag it in the recap as a decision.
[Marcus] I will have the brief ready Thursday morning.
[Priya] And let us pull the Q3 numbers first.
Double-tap ★ bookmark ↓ end
It works the way you do.
Brief me before the call.
How it works
Numbered, ranked, on the lens. Walk in already knowing the three things that matter, read straight from your meetings and tasks.
Say it once, hands-free.
How it works
Dictate a long prompt or reply by voice. With on-phone Moonshine, text forms and revises while you speak, then the existing finalizer cleans the settled draft. Review it, edit by voice, or re-record before a single word is sent.
Every meeting, captured.
How it works
Fast Moonshine preview can appear while you speak, then speaker-attributed Large-v3 replaces it as the canonical transcript. Takeaways and action items wait for you afterward, right from the glasses.
Record a message
Say it once. Send it clean.
Dictate a long prompt or reply by voice. Moonshine Tiny or Base can form and revise the text on the lens as you speak; the shared finalizer cleans the settled draft before review. No keyboard required.
MESSAGE · COS [O●] 100%
Listening…
Tap to finish
On the lens
In your control.
Confirm, edit, or send with a tap. The HUD is there when you need it and gone when you do not. Never a phone, never a screen between you and the room.
