Hinty
Get Started
← Hinty Blog
AI & Job Interviews

Claude Code on Smart Glasses: Approve Agent Steps Without Your Desk

Hinty TeamNovember 21, 202628 views
Claude Code on Smart Glasses: Approve Agent Steps Without Your Desk
A coding agent works for ten minutes, hits a step it is not allowed to take alone, such as running your tests or deleting a folder, and then it waits. If you are not at your desk, it waits until you come back. Claude Code can run on smart glasses today in the sense that matters: you can see that an agent is waiting and approve or deny the step from your face, your phone or your wrist. As of November 21, 2026, the glasses route is Even Realities Terminal Mode on the G2, plus a few community projects.

Hinty does not control Claude Code or any other agent today. What Hinty Voice Assistant does on glasses today is live hints in a conversation: when someone asks you a question, a short answer is spoken in your ear, or shown in the lens on glasses with a display. Agent supervision is something we are exploring, and this post says plainly where it stands.

Why agents wait for you

A coding agent asks before it does anything with side effects. Claude Code asks before running a shell command it has no rule for, and Codex has similar approval gates. That is the right design, and it is also why a long job can sit idle for an hour while you are in a meeting or on a walk.

Two things make the wait expensive. You often do not know the agent is waiting, because the only signal is a terminal you are not looking at. And the question is tiny: "Run npm run test:e2e?" takes two seconds to answer once you see it. Glasses fit both well, with a few words in your field of view and a gesture back. The part most demos skip is that a tap approving a command on your computer is a remote control for your shell. That is where the options differ.

What you can use today

There are four kinds of options: Even's own Terminal Mode, community projects for glasses, official phone apps from the agent makers, and watch projects. Everything below was checked on public pages on November 21, 2026. Items we could not confirm are left out.

OptionDeviceAgentsNeedsHonest limit
Even Terminal ModeEven G2, R1 ringClaude Code, CodexComputer running the tool, same network or a tunnelRing listed as required in the readme; each session is its own
Claude Code Remote ControlPhone, browserClaude CodePro, Max, Ultra, Team or Enterprise plan, no API keysLocal process must keep running; one agent per session
Codex on ChatGPT mobilePhoneCodexCodex app on a Mac (preview)Setup is through the desktop app, not the CLI
cc-g2 and similarEven G2Claude Code, CodexMac, iPhone, TailscaleExperiment; the author says so
claude-watchApple WatchClaude CodeMac, iPhone, Xcode, same Wi-FiCommunity project; traffic encryption not described

Even Terminal Mode on the G2

This is the closest thing to an official answer. Even's Terminal page says it runs your coding agents on your own machine and shows "what is running, blocked, or waiting for input across multiple sessions" on the G2. You approve simple actions with a ring tap, and you tap and hold to dictate a short instruction (Even Realities, checked November 21, 2026).

The npm readme for the tool, version 0.10.5 (published September 24, 2026), fills in the details: it needs Node.js 18 or later and lists an Even G2 plus an R1 ring paired through the Even app. It supports Claude Code and Codex, and it runs your agent through its own process. By default it listens on your local network. To reach it from outside, it suggests Tailscale or a temporary public tunnel such as ngrok, pinggy or bore (npm, registry data read November 21, 2026). Even's own page says reaching the host from outside your Wi-Fi needs "advanced setup".

The readme also notes that some of Claude's default tools, such as web search and web fetch, are auto-approved without an app confirmation.

It launched on Product Hunt in June 2026 and ranked fourth for June 11 with 336 upvotes (Product Hunt). One commenter there asked how it avoids approval fatigue from repeated "allow once" prompts, and the maker's answer did not address it directly. That question comes back in the security section below.

Remote Control and the Claude app

If you do not own glasses, Remote Control is the simplest option. It connects the Claude mobile app or claude.ai/code to a Claude Code session running on your machine. Execution and file access stay local. It works on Pro, Max, Ultra, Team and Enterprise plans, not with API keys, and not on Bedrock, Vertex or Foundry setups (Claude Code docs, read November 21, 2026).

Two limits matter here. The local process must keep running: if you close the terminal or quit the app that hosts the session, it goes offline. And while connected, the session transcript, including tool activity, is stored on Anthropic's servers. Claude Code keeps permission prompts open until you answer them, and the docs describe push notifications for actions that need you.

Codex on your phone

OpenAI announced Codex in the ChatGPT mobile app on May 14, 2026. From your phone you can answer Codex's questions, approve commands and review diffs. It was a preview for all plans and needed the Codex app on macOS, with Windows "coming soon" at that time. Files, credentials and permissions stay on the machine where Codex runs (OpenAI). We did not confirm the current Windows status.

Community projects on glasses and wrists

Several individuals have built their own bridges, and they are honest about their limits:

  • cc-g2 lets you approve or deny Claude Code and Codex requests from the G2. A comment you add on an approval screen is forced into "deny plus comment", so the agent asks about intent instead of proceeding. The author's setup uses Tailscale, calls it an experiment and says it is not a replacement for Even's official tool (Zenn, March 9, 2026, updated May 3, 2026).

  • claude-code-g2 runs Claude Code by voice from the G2. Its readme warns that the access token can run shell commands on your Mac, and that Claude runs with permission prompts skipped by default (GitHub, MIT).

  • claude-watch puts permission prompts on an Apple Watch through a bridge on your Mac. It needs Xcode 16 or later to build, the watch on the same Wi-Fi as the Mac, and its readme does not describe how traffic is encrypted (GitHub, MIT, 645 stars, read November 21, 2026).

  • OcuClaw is a separate platform for the G2 that runs Hermes Agent and OpenClaw, not Claude Code. Its page says it is not affiliated with Even Realities and does not list a price (ocuclaw.com).

  • Monako Glass is a coding-focused pair announced in June 2026. Digital Trends reported on June 4, 2026 that it names Claude Code and Codex but did not explain how they run, and that no price or shipping date was posted (Digital Trends). Treat it as an announcement.
  • For a wider survey, see AI agents on smart glasses. For the other direction, building glasses apps with coding agents, see what agents get wrong about smart glasses.

    Try it in 2 minutes: live hints on your glasses, today

    This section is about what Hinty does today, which is not agent control. It works on glasses you may already own.

    Scene: an interviewer says, "Tell me about a time you missed a deadline." About two seconds later, in your ear: "Q3 launch, two weeks late, what you changed." On Ray-Ban Meta and Oakley Meta the hint is spoken, not shown, because those glasses have no display. Rokid Glasses show it in the lens. Nothing is installed on the glasses, and the same works with any Bluetooth earbuds.

  • Install Hinty Voice Assistant on your phone (Google Play or App Store) and sign in.

  • Pair your glasses or earbuds as normal Bluetooth audio.

  • Open the glasses screen in the app and send a test hint.

  • Start a conversation and let your phone listen.
  • Step-by-step pages with store buttons are in the setup guides. If you want to be told first when agent supervision reaches testers, sign up on the smart glasses page.

    Price, as of November 21, 2026: free to start. The Interview Pass is 9.99 USD one-time on hinty.eu for 120 minutes, with no subscription and minutes that do not expire. Monthly plans are 7.50, 10 and 20 USD for 45, 90 and 180 minutes.

    What goes wrong when you approve from one line of text

    A glasses display shows a few short lines, so a long command may be shown only in part. Here is what can go wrong, and none of it needs a malicious tool.

  • The visible part is harmless, the rest is not. A command like "npm test" followed by a long chain that fetches and runs something else looks fine in the first line. If you can only see the beginning, you approve the beginning.

  • A summary written by the agent. Many setups show a short line the agent wrote about its own step. A prompt injection hidden in a file the agent read can make that line sound routine. Summary and command should be shown separately.

  • Approval fatigue. Forty "allow once" prompts a day teaches your thumb to say yes. This is the exact question raised on Product Hunt.

  • An accidental tap. Adjusting glasses on your nose should never run a command.

  • A permanent yes. An "always allow" choice from a remote device widens permissions for every later session.

  • Silence read as consent. If the connection drops, the agent should ask at the terminal, not proceed.
  • This is why we think two-step consent matters. In the design we are exploring, a single tap never approves anything: you scroll to "Allow once" and then tap, so two deliberate gestures are needed. Only "allow once" and "deny" are offered from the glasses, never "always allow" and never an edited command. If the command is cut off on the display, the glasses can deny it, but the approval has to happen on a screen that shows the whole text. No answer means no decision. These rules cost a few seconds per approval, and that is the point.

    What Hinty is building

    Hinty Agents is in closed testing and is not public. There is no date, and testers hear first. It is a direction, not something you can install:

  • see several agents at once, in one list, instead of one terminal per session;

  • approve or deny a waiting step with two gestures, as described above;

  • say the next instruction by voice;

  • work from hooks on sessions you already run, without a separate setup for each agent and without a private network.
  • The difference from Even's own tool, as we see it, is sessions you already run, one list across computers, and glasses other than the G2. Whether that is worth building depends on what testers tell us. Tester sign-up is on the smart glasses page, and the test is free.

    Who this is not for

    If you want Claude Code on your glasses this week, Even Terminal Mode with a G2 and a ring is the most direct route we found, and it needs a computer that stays on and some terminal comfort. If you do not own glasses, Remote Control on your phone is cheaper and officially supported. Hinty does not do agent control today, so do not buy anything for that reason.

    Hinty's live hints need your phone nearby and internet. On glasses without a display, hints are voice only. If you use hints in a conversation, tell the other person, and check local recording and consent laws.

    FAQ

    Can Claude Code run on smart glasses?

    Not on the glasses themselves. Claude Code runs on your computer, and the glasses act as a screen and remote. Even Terminal Mode does this on the G2, and community projects do the same for Claude Code and Codex.

    Do I need the R1 ring for Even Terminal Mode?

    The npm readme lists a G2 and an R1 ring as requirements, and Even's page describes approving with a ring tap. Some secondary sources call the ring optional. Check Even's current instructions before you buy.

    Can I approve Claude Code permissions from my phone?

    Yes. Remote Control connects the Claude mobile app to a session on your computer. It needs a Pro, Max, Ultra, Team or Enterprise plan, and the session must keep running on your machine.

    Is it safe to approve a command from glasses?

    It is as safe as the rules around it. Keep the agent's permission rules on, use a private network or tunnel you trust, and prefer setups where silence means no and a single tap cannot approve.

    Does Hinty work with Claude Code?

    No. Hinty does not control Claude Code or any other agent today. We are exploring agent supervision from glasses, with no date, and testers hear first.

    Is there a free version of Hinty?

    Yes, you can start free. The Interview Pass is 9.99 USD for 120 minutes, and monthly plans are 7.50, 10 and 20 USD, as of November 21, 2026.

    Hinty is not affiliated with Even Realities, Anthropic, OpenAI, Meta or any glasses maker named here. Names describe compatibility only.

    Next step: pick your glasses and open the setup guide for Ray-Ban Meta or the Bluetooth headset guide.


    Keep reading

  • AI agents on smart glasses: what actually works in 2026

  • Coding agents and smart glasses: 39% right from memory, 89% with skills

  • Smart glasses apps: which glasses let you install them

  • Hear first about agents on glasses

    Hinty already puts live hints in your ear or lens. Join the testers to try new glasses features first, including any agent supervision we build.

  • First access to every new glasses feature

  • Free during the test

  • Your feedback decides what we build next
  • 👉 Become a tester

    👉 Set up Hinty on your glasses

    #claude code smart glasses#coding agents glasses#even realities terminal mode#claude code remote control#approve ai agent permissions#smart glasses for developers

    Hints in your glasses · free

    Try it

    Related articles

    Rate this article

    Average: 0.0 (0 ratings)

    Login to rate this article

    Comments (0)

    Login to add a comment

    No comments yet. Be the first!