From your glasses to your spreadsheet

Five things. With all five, nothing in part one can really go wrong.

• Even G2 glasses, paired and working with the Even Realities app on your phone.
• The Even Realities app, updated. An out-of-date Even app is the single most common reason part one half-works.
• IRIS installed from the Even Hub store.
• IRIS Pro. Quick capture and the agent bridge are both Pro features, and the free trial includes them.
• For parts two and three only: a computer that can stay switched on, because that is where your agent lives.

No computer, no terminal, nothing to install. Ten minutes on your phone, and afterwards you can talk to IRIS with the app closed and the phone in your pocket.

What you get

Say "Hey Even, remember to send the deck" and it becomes a to-do. "Hey Even, note that the gate code is 4821" becomes a note. "Hey Even, remind me at 9 to take out the trash" becomes a timed reminder, emailed to you at that moment. "Hey Even, put lunch with Sam on my calendar Friday at noon" becomes a card with an Add to calendar button. And you can ask it things — "Hey Even, what did Diane say?" is answered from your own IRIS memory, on the lens.

The setup

1. Open IRIS and find the two values

   Go to Settings → Connections → Even AI quick capture. With a licence key entered, you will see an Agent URL and a Token, each with a Copy button.

   Agent URL

   https://iris.eveniris.com/agent

   Token

   your IRIS licence key

   Copy the URL now. You will come back for the token in a moment — most phones hold only one thing on the clipboard at a time.
2. Add IRIS as an agent in the Even app

   In the Even Realities app: Settings → Even AI → Agent Configuration → Add Agent. Name it IRIS, and paste the Agent URL into the URL field.
3. Go back for the token

   Switch to IRIS, tap Copy next to the Token, switch back to the Even app, paste it into the token field, and save.
4. Select IRIS — this is the step people miss

   Adding an agent is not the same as using it. In Even's Settings → Even AI, make sure IRIS is the selected agent. If something else is still selected, your words go there instead and nothing arrives.
5. Test it

   With IRIS closed, say "Hey Even, remember to buy milk". Open IRIS and look in Memory → Kept — your capture is there. Then try a question, "Hey Even, what am I meant to buy?", and the answer appears on the lens in a second or two.

Capture and memory recall are dependable. Live web questions through this path — news, weather, scores — are still in beta and can be slow or unreliable, which is worth knowing before you lean on it in front of someone.

IRIS can hand your words to an agent you run yourself and stream its reply back to your lens. This is the part that later lets it touch your files. Everything here happens in Settings → Connections → Your agent.

The distinction that decides everything

Which one do you want?

How Claude Code gets onto your glasses

The arrangement is IRIS → OpenClaw → Claude Code → your files. OpenClaw runs command-line agents as backends, and its bundled Anthropic plugin registers Claude Code as the claude-cli backend, driving the claude executable you already have installed.

1. Get Claude Code working in your terminal first

   If typing claude does not start it, nothing downstream will work. Fix that before going on.
2. Turn on the endpoint IRIS speaks to

   OpenClaw's OpenAI-compatible endpoint is off until you ask for it. In your gateway config:

   gateway: { http: { endpoints: { chatCompletions: { enabled: true } } } }

   Restart the gateway.
3. Point your agent at the Claude Code backend

   In OpenClaw's agent configuration, set the model to a claude-cli/ reference. The exact form moves between versions, so take it from OpenClaw's CLI backends documentation rather than from memory.
4. Give the gateway a public address

   The gateway listens on port 18789. In a terminal:

   cloudflared tunnel --url localhost:18789

   It prints an address ending trycloudflare.com. Leave that window open — closing it closes the tunnel.
5. Fill it in on the phone

   In IRIS: Settings → Connections → Your agent → OpenClaw. Paste the tunnel address into OpenClaw gateway URL, and your gateway token — gateway.auth.token in the config, or the OPENCLAW_GATEWAY_TOKEN variable — into Gateway token. Leave Agent ID blank to talk to your default agent. Tap Test connection.

Or just a model, in two minutes

Tap Other model and fill in three things: the base URL, your key, and the model name. The base URL is the one the provider's documentation ends with /v1 — the trailing /v1 is optional here, it works with or without.

The model name is not optional and has no sensible default — a provider rejects every model id but its own. Copy it character for character from their documentation. Until it is filled in, IRIS deliberately shows the agent as not connected, because a guessed name only fails later as a confusing error.

What being connected gets you

• The Ask tab gains a toggle between IRIS and your agent.
• A Terminal tile on Home: type a command, your agent streams a result back, and you can push it to the glasses. It asks you to tap twice before sending, because the text reaches a real machine.
• During a live session, a button sends the conversation to your agent and the reply lands on your lens.
• Auto-send sessions to agent memory pushes each session summary into a real agent's long-term memory as it finishes. Not offered for a plain model, which has no memory to push into.
• Let your agent write to your lens, off by default. The only thing that can put words on your glasses unasked: a reminder, the result of a long job, appearing as a cue during a session. Anything sent outside a session waits and lands in your next one.
• More than one agent. The first is your default; each mode can point at a different one, so work and personal can run on separate agents.

Hold onto this: IRIS is the voice and the display. She never touches your files herself. Everything in this part is configured on your agent's machine — and once it is done, it is available from the glasses automatically, with nothing further to set in the app.

Route A: files on your own computer

The simplest thing that works, and the one to try first. An agent like Claude Code already reads and writes files in a folder you give it, including .xlsx workbooks and .docx documents, which it handles by writing and running small scripts against them.

1. Make one folder and put the files in it

   Something like ~/agent-files. One folder keeps the blast radius small and makes it obvious what your agent can see.
2. Make that folder sync, if you want the changes online

   Install Google Drive for desktop or OneDrive and keep the folder inside the synced area. Your agent edits the local file; the sync client pushes it up within seconds.
3. Say what you want, into your glasses

   "Open the Q3 budget workbook and add a row for the new supplier, three thousand four hundred, dated today."

Route B: connect the agent to real services

Agents reach outside tools through MCP — a standard way of plugging a capability into an agent. You install a small server for the thing you want to control, register it with your agent once, and from then on it is simply another thing your agent can do.

Excel, without Excel installed

The Excel MCP server creates and edits workbooks — data, formatting, formulas, charts, pivot tables — and needs no copy of Microsoft Excel on the machine. To register it with Claude Code:

claude mcp add --transport stdio excel -- uvx excel-mcp-server stdio

Google Docs, Sheets, Drive, Gmail, Calendar

The Google Workspace MCP server covers the Google side properly — Docs editing, Sheets ranges and formatting, Drive, Calendar, Gmail, Tasks. It needs a Google OAuth client, which is the same one-time Google Cloud setup IRIS asks for when you switch on Drive exports.

export GOOGLE_OAUTH_CLIENT_ID="your-client-id"
export GOOGLE_OAUTH_CLIENT_SECRET="your-client-secret"
claude mcp add --transport stdio gworkspace -- uvx workspace-mcp --tool-tier core

Google also runs an official remote Drive server at https://drivemcp.googleapis.com/mcp/v1. It is the first-party option, but note what it actually does: searching, reading, creating and copying files. Editing the contents of an existing Doc is not among its tools.

Checking it took

claude mcp list

Inside a Claude Code session, /mcp shows the connected servers and runs the browser sign-in for any that need one. If your agent is OpenClaw, it bridges MCP tools through to its command-line backends with bundleMcp: true, so the tools you add reach Claude Code underneath.

What it actually feels like

Once route A or B is in place, this is the sort of thing you say — through the Ask tab on your glasses, or typed into the Terminal tile.

You are about to let something act on your files from a pair of glasses. Four habits make that comfortable rather than nerve-racking.

1. Turn on approvals

   With Let your agent write to your lens switched on, your agent can ask you to approve an action before taking it. A one-line summary appears on the lens; a tap approves, a swipe or double-tap refuses. Refusal is the default everywhere — if it expires, if you never see it, if the app is closed, the answer is no. An answer once given cannot be changed.
2. Start read-only

   Spend the first day asking questions about your files before letting anything write to them. You learn what it understands, cheaply.
3. One folder, backed up

   Give it a working folder rather than your whole home directory or your entire Drive. Keep a backup, or rely on version history. The point is that a mistake is recoverable, not that it is impossible.
4. Keys are passwords

   Your gateway token, your API_SERVER_KEY, your provider key. IRIS passes them through per request and stores none of them, but treat them the way you would treat the password to the machine itself.

The messages are specific on purpose. Each one means one thing. For the agent bridge in particular, the agent page has a longer list.

Every term used above, in the order you meet it.

Hey Even

The Even Realities voice assistant built into your glasses. It can be pointed at an agent of your choosing, which is what part one does — pointing it at IRIS.

Licence key

Your IRIS Pro key, entered in IRIS Settings. It is also the token Even's agent configuration asks for.

API key

A long secret string from a service that proves a request is yours. Treat every one of them as a password.

Base URL

The web address a provider's API lives at, usually ending in /v1. Their documentation always states it.

Model

The thing that produces answers. No memory of its own, no tools, no access to your machine.

Agent

A program on a machine you control, with memory and tools, that can actually do things. Hermes, OpenClaw, Claude Code.

Claude Code

A command-line agent. It does not listen for incoming connections, so IRIS reaches it through OpenClaw rather than directly.

Gateway

The part of OpenClaw that listens for requests and hands them to whichever agent should answer.

Tunnel

A small program that gives a service on your computer a public web address, so your phone can reach it from anywhere. cloudflared is the one used here.

MCP

The standard way of plugging a capability — Excel, Google Drive, a database — into an agent. Install a small server once and your agent can use it from then on.