IRIS

A real-time AI assistant for Even Realities G2 smart glasses. It listens to your conversations and surfaces discreet cues on your lens — answers, follow-up questions, fact-checks, and more — exactly when they're useful.

What can it do?

IRIS works entirely in the background while you talk. You don't need to look at your phone or interrupt the conversation. Cues appear on the lens in a line or two, glanceable in under a second.

Everyday conversations

Get quick answers when someone asks you something, gentle fact-checks when something sounds off, or a good follow-up question to keep things going.

Listening & support

Warm prompts to help you be a better listener — specific follow-ups that show you were paying attention, or supportive suggestions when someone's struggling.

Lectures & talks

Quietly define technical terms as they come up, identify people or works mentioned, or answer questions the speaker poses to the room.

Your own modes

Describe any situation — job interviews, parent–teacher meetings, sales calls, language practice — and IRIS generates a mode tailored to it.

Setup

IRIS is install-it-yourself — you bring your own AI and speech keys. There's no subscription and no data goes to IRIS servers. Setup takes about five minutes.

1

Install from Even Hub

Open the Even Realities app on your phone, go to Hub → search "IRIS", and install.

2

Choose an AI engine

Pick the provider you want to use for cue generation. Groq and Gemini both have free tiers with no card required — good starting points. See AI engine for the full comparison.

3

Choose a speech provider

Pick who transcribes your speech. Deepgram is the default and the most accurate for English. Free tiers are available. See STT providers.

4

Enter your API keys

Paste in the keys for your chosen providers in Settings. Keys are stored only on your device and sent directly to those services — nothing passes through IRIS.

5

Start a session

Tap the glasses touchpad once to bring up the start prompt, then tap again to confirm. Or hit the big button in the IRIS app. That's it.

If you're using Claude (Anthropic): Claude routes through a personal relay (a small Cloudflare Worker) to keep your key off the phone. This takes an extra 10-minute setup step — full instructions are in the app after you select Claude.

AI engine

The AI engine decides what cue to show you and when to stay silent. You can switch providers any time from Settings → AI engine — your keys for other providers are always saved and ready to switch back to.

Paid

Claude Haiku 4.5 (Anthropic)

The most accurate at knowing when not to interrupt. Excellent cue quality and strong language support. Costs a fraction of a cent per conversation — you prepay a few dollars of credit at console.anthropic.com. Requires the relay setup step.

Paid

GPT-4o (OpenAI)

Excellent all-round quality and reasoning, broad language coverage, and connects directly from the phone — no relay needed. Slightly more eager to suggest cues than Claude. Paid only.

Free tier

Gemini 2.5 Flash (Google)

The most generous free tier of the lot, no card required. Very fast and covers a huge range of languages. A notch below Claude/GPT on subtle judgment, but great as a zero-cost starting point.

Free tier

Groq (Llama 3.3 70B)

Fastest response times and a free tier, no card needed. Best for budget or backup use. Lower cue quality — can over-trigger fact-checks — but fine for basic conversation assist.

Your setup

Custom (OpenAI-compatible)

Point IRIS at any OpenAI-compatible API — local models like Ollama or LM Studio, or any third-party host. See Local / private models.

Local / private models

If you work with sensitive data and don't want conversations leaving your network, you can run a local language model and point IRIS at it. Nothing leaves your device or LAN.

How to set it up

1

Install Ollama or LM Studio on your computer

Both run local models and expose an OpenAI-compatible API on your LAN. Download from their respective sites.

2

Make the server listen on the network (not just localhost)

For Ollama: run OLLAMA_HOST=0.0.0.0 ollama serve. For LM Studio: enable "Server" mode and set it to listen on all interfaces.

3

Find your computer's local IP address

On Mac: System Settings → Network. On Windows: ipconfig in Terminal. It'll look like 192.168.1.42.

4

Enter the details in IRIS Settings

Go to Settings → AI engine → Custom. Enter the base URL (e.g. http://192.168.1.42:11434/v1 for Ollama), your model name (e.g. llama3 or qwen2.5), and leave the API key blank.

Phone and computer must be on the same Wi-Fi network. This won't work over mobile data. Switching back to a cloud provider takes one tap in Settings — all your other keys are still saved.

Recommended models for IRIS

IRIS sends short prompts and needs the model to return a single JSON object. Most capable 7B+ instruction-tuned models work. Good choices: Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B Instruct, Phi-4. Bigger models give better cue quality but are slower.

Speech recognition (STT)

The speech provider turns what you and others say into text. IRIS sends that transcript to the AI engine to decide on a cue. You can switch STT providers independently from your AI engine — your keys for all providers are always saved.

Free tier

Deepgram (default)

Best accuracy for English and the widest support for other major languages. The default choice. Connects directly via WebSocket for low-latency transcription. Free tier available at console.deepgram.com.

Paid

Soniox

Strong multilingual accuracy, especially for languages Deepgram struggles with. Good fallback if you're working in a less common language. Direct WebSocket connection.

Paid

OpenAI Realtime

Uses OpenAI's Realtime transcription API. Good accuracy in many languages. Uses your OpenAI API key — no separate key needed if you already have one for GPT.

Free tier

Gemini

Google's transcription via the Gemini API. Reasonable accuracy and very broad language coverage. Uses your Gemini API key. Good pairing with the Gemini AI engine.

Language selection: Set the language in Settings → Language. "Auto" lets the provider detect the language automatically — covers the major ones well. If you speak a less common language, pin it explicitly for better accuracy.

Modes

A mode tells IRIS what situation you're in and how it should behave — what kinds of cues to give, how often, and in what tone. You can have as many modes as you like and switch between them with a scroll on the glasses.

Built-in modes

Conversation

General back-and-forth. Surfaces answers to factual questions, gentle corrections when something's wrong, and good follow-up questions. Stays quiet during small talk.

Listening

For when someone else is talking and you want to be fully present. Offers warm follow-up questions and supportive suggestions. Avoids generic prompts.

Study

For lectures and presentations. Defines terms, identifies people or works mentioned. Never tells you to speak — you're quietly absorbing. Includes periodic recaps.

Creating a custom mode

Go to the Ask IRIS tab and describe what you want. For example: "Make me a mode for job interviews where I'm the candidate" or "I need a mode for my weekly team stand-ups." IRIS generates a mode from your description and asks if you'd like to apply it.

You can also edit any mode's name, description, and settings manually from Settings → Modes.

Tuning a mode

Each mode has a few knobs to adjust:

Chattiness

Low Very sparing — only surfaces a cue when it's clearly valuable. Default to staying silent. Best for formal or high-stakes conversations where you don't want unexpected interruptions.
Medium The default. Surfaces a cue when it's genuinely useful, otherwise stays quiet. Good balance for most situations.
High Leans toward being helpful — suggests a cue whenever there's a reasonable one. Good for fast-paced conversations or when you want IRIS to be more proactive.

Silence prompts

When the conversation goes quiet for 30 seconds or more, IRIS can offer a gentle prompt to get things going again — a question you could ask, something worth sharing. Enabled by default. Turn it off for modes where silence is fine (e.g. listening to a lecture).

Recaps

For long sessions like lectures, IRIS can periodically surface a brief summary of what's been covered so far — useful to help you keep the thread. Off by default; enable per mode.

Cue types

Choose which types of cues this mode can produce. See Cue types for what each one means. Restricting a mode to only the relevant types (e.g. only "follow-up" for a listening mode) keeps cues focused.

Live sessions

A session is an active period of listening. IRIS transcribes everything in real time, feeds it to the AI engine, and shows cues on your lens as they arrive. Your transcript and cues are saved automatically.

Starting a session

Tap the Start button in the IRIS app, or use the glasses touchpad (see Glasses controls). The app can be in the background — once a session starts, it keeps running with the phone screen off.

During a session

On-lens transcript (optional)

Enable On-lens transcript in Settings to show a rolling strip of the live transcript along the bottom of the lens — useful if you want to confirm IRIS is hearing correctly. Cues appear in the upper area as usual.

Ending a session

Tap End in the IRIS app. The session is saved and IRIS generates a title and summary in the background. This takes under a minute.

Telling IRIS to be quiet

During a session, type something like "don't give me any cues for now" or "stop responding" in the context panel and IRIS will pause. Type anything else to resume, or tap Resume cues.

Glasses controls

The G2 touchpad on the right temple controls IRIS entirely from the glasses — no need to touch your phone during a conversation.

When idle (not in a session)

Single tap Opens the "Start [Mode Name]?" confirmation screen.
Tap again Confirms and starts the session.
Scroll up / down Cycles through your available modes (shown on the lens).
Double tap Exits the IRIS HUD (shows the OS exit confirmation).

During a session (recording)

Scroll down Opens cue history — scroll up/down to browse previous cues.
Single tap (in cue history) Returns to the live recording view.

When a cue is showing

Single tap Dismisses the cue and returns to the recording view.
Sessions end from the phone only — there's no glasses gesture to stop recording. This is intentional: it prevents accidental session endings when you're scrolling through cues.

Cue types

Each cue is tagged with a type that tells you at a glance what it's for. The lens shows the tag abbreviation; the app shows the full label.

Answer [ANS] A direct answer to a factual question someone just asked. Read it out or adapt it naturally.
Follow-up [ASK?] A question you could ask the other person to deepen the conversation or show genuine interest.
Explanation [INFO] Background context, a definition, or clarification — useful when a term or concept came up that's worth understanding.
Fact-check [!] Flags when something the other person said is likely factually wrong. Only surfaces when IRIS is highly confident — not for opinions.
Advice [TRY] A concrete suggestion for something you could say or do, especially when someone is asking for help or guidance.
Thought [SHARE] An insight, observation, or point worth sharing — for modes where you want IRIS to prompt you to contribute ideas.

Prep notes

Before starting a session, you can type context into the Prep notes field on the home screen. IRIS treats this as instructions for that session — it's injected directly into the prompt alongside the transcript.

Examples of what to put there:

Prep notes are saved with the session so you can refer back to what you briefed IRIS on.

Knowledge base

The knowledge base is a text block you can fill with reference material you want IRIS to draw on — facts, names, terminology, background on people you talk to, anything useful. It's attached to specific modes, not to sessions.

When a mode has relevant document content set, IRIS automatically refers to it when generating cues. Useful for:

Set it in Settings → Modes → [your mode] → Reference material.

Session history

Every session is saved automatically with its full transcript, cue history, and a generated title and summary. Find them in the Sessions tab.

After a session ends, IRIS runs a few background steps:

If the app is closed before processing finishes, it picks up where it left off the next time you open it.

Memory

IRIS builds a long-term picture of you across sessions — things you care about, people in your life, recurring topics, preferences. This context is quietly included when generating cues, making them more relevant over time.

You can see, edit, and delete memory entries in the Memory tab. Nothing is shared outside your device and your chosen AI provider.

Weekly review

At the end of each week, IRIS can generate a synthesis across all your sessions — patterns, recurring themes, progress on goals, and things worth reflecting on. Find it in the Sessions tab under "Weekly review".

It's generated on demand, not automatically — tap Generate to run it whenever you're ready.

Tips & tricks

Start with Groq or Gemini

Both have free tiers and need no card. They're a good way to try IRIS without any commitment — you can switch to Claude or GPT later once you've got a feel for it.

Use chattiness to tune the signal-to-noise ratio

If you're getting too many cues, drop to Low. If IRIS feels too quiet, switch to High. This is per-mode, so you can have a quiet mode for sensitive conversations and a chatty one for casual use.

Tell IRIS what's happening mid-session

The context panel (in-app, during a session) lets you type instructions at any time: "The person I'm talking to is a doctor", "We just moved to talking about salary", "Stop giving me cues for a bit". IRIS picks these up immediately.

Keep the app open while you set up

All keys are stored on your device. If the app closes mid-setup, nothing is lost — just reopen it and continue.

Sessions keep running with the screen off

Once a session starts, you can pocket your phone. IRIS keeps recording and generating cues on the glasses until you tap End. The phone doesn't need to be visible.

For local models: use a 7B+ model

Smaller models (1B–3B) often struggle to follow the JSON output format IRIS requires. A 7B+ instruction-tuned model is more reliable. Qwen 2.5 7B and Llama 3.1 8B are good starting points.

Ask IRIS to make your mode better

In the Ask IRIS tab, you can say things like "My cues are too generic — make them more specific to technical conversations" or "I'm getting too many fact-checks, tone it down" and IRIS will update your active mode accordingly.