Cluely’s Rival Halo Ships Meeting Copilot 2026: Real Test

Halo, the invisible desktop assistant from Stanford dropouts Patrick Shen and Ethan Zhang, just shipped the pieces that turn it from a demo into a product: an open-source client called Glass and a paid team tier pointed directly at Cluely. The story isn’t the funding-drama subplot between two rival founders — it’s the architecture. A Halo AI meeting assistant listens at the operating system’s audio and screen layer, so it never joins your Zoom, Teams, or Meet call as a participant. That single design choice sidesteps the wave of enterprise IT policies now blocking bot notetakers outright, and it drops a consent problem in your lap that no vendor has cleanly solved.

What’s new with the Halo AI meeting assistant

Halo (shipped under the Pickle umbrella, with the client named Glass) is a desktop overlay that captures system audio and reads your screen, then feeds both into a live model context. It renders answers in a floating window excluded from screen-capture APIs on macOS and Windows. The window doesn’t appear when you share your screen, and it doesn’t appear in the recording. There’s no bot in the participant list because there’s no bot. The app is a local listener, not a guest.

Two shipped items matter. First, the Glass client is open source, so you can read the capture code, self-host the backend, and swap in your own model keys instead of routing audio transcripts through a vendor. That’s a genuine differentiator against Cluely, whose stack is closed. Second, the paid team tier adds shared prompt libraries, org-level context (product docs, objection-handling scripts, pricing sheets), and admin controls — the boring enterprise plumbing that separates a viral demo from a line item on a sales team’s budget.

The competitive framing is deliberate. Cluely spent 2025 building a brand on “cheat on everything,” and it worked as marketing right up until enterprises started reading it as a threat model. Halo’s positioning is the same technology with a compliance-shaped face on it: open source client, self-hostable, no bot in the room. Whether that reframing survives contact with actual legal departments is the open question, and the one worth watching.

Why it matters

• Bot blocking is now a real deployment blocker. A growing share of enterprise Zoom and Teams tenants disable third-party meeting bots at the admin level. Any product whose entire capture strategy is “join the call as a participant” — Otter, Fireflies, Fathom, Read — sits one policy toggle away from being unusable inside your biggest accounts. An AI meeting copilot without a bot has no such dependency.
• Open source changes the procurement conversation. “Where does our audio go?” is the first question security review asks. An open source meeting notetaker you can self-host answers it without requiring trust in a vendor’s SOC 2 PDF. That’s a structural advantage over closed competitors, not a marketing one.
• Consent moves from the platform to you. When a bot joins, everyone sees “Notetaker has joined” and the platform arguably handles disclosure. When an invisible AI assistant listens from your desktop, disclosure falls entirely on you — and in two-party-consent jurisdictions (California, Florida, Pennsylvania, Illinois, Washington, and others), recording without it is a criminal matter, not a policy violation.
• The real-time interview assistant use case is a reputational landmine. Halo’s screen-invisible overlay straightforwardly feeds answers during a live technical interview. Employers are already responding — some are moving back to in-person loops specifically because of this class of tool. Expect that arms race to escalate.
• The notetaker category is being unbundled. Transcription is a commodity; Whisper-class models run locally for free. The remaining value sits in real-time assistance during the call and structured extraction after it. Halo attacks the first, most incumbents own the second, and the two will collide.
• Screen context beats audio-only context. A tool that reads the shared deck, the spreadsheet, and the error stack on screen has strictly more signal than one working from a transcript. That’s the durable technical argument for the screen-and-audio-layer approach, independent of the bot politics.

How to use the Halo AI meeting assistant today

1. Clone and build the Glass client. The open source client is the honest starting point — read exactly what it captures before you let it capture anything.

   git clone https://github.com/pickle-com/glass.git
   cd glass
   npm install
   npm run build:all
   npm start

2. Bring your own model key instead of the hosted backend. Self-hosting keeps transcripts on infrastructure you control. Set your provider key in the local environment before first launch:

   # .env in the project root
   ANTHROPIC_API_KEY=sk-ant-...
   # or, if you prefer a local model endpoint:
   OPENAI_API_BASE=http://localhost:11434/v1
   OPENAI_API_KEY=ollama

3. Grant only the permissions you need. On macOS the app needs Screen Recording and Microphone under System Settings → Privacy & Security. System audio capture on macOS requires a loopback driver or the ScreenCaptureKit audio path — verify which one your build uses. On Windows, WASAPI loopback handles it natively. Check what’s actually granted:

   # macOS — list apps holding Screen Recording permission
   sqlite3 ~/Library/Application\ Support/com.apple.TCC/TCC.db \
     "SELECT client, auth_value FROM access WHERE service='kTCCServiceScreenCapture';"

4. Write a system prompt that fits the meeting type. The default “answer everything” behavior produces noise. Constrain it. A sales-call prompt that has worked well for us:

   You are a live sales copilot. You see my screen and hear both sides of the call.
   
   Rules:
   - Stay silent unless one of these fires:
     1. The prospect raises a pricing or competitor objection.
     2. They ask a factual question about our product I may not know.
     3. They state a number, date, or commitment worth capturing.
   - Output max 20 words. No preamble. No restating their question.
   - For objections: give the counter, not the theory behind it.
   - Cite the source doc name in brackets when you pull from context.
   
   Context: [paste pricing sheet, objection handling doc, product one-pager]

5. Disclose before you listen. Non-negotiable, and cheap to automate. Put this in your calendar invite template and say the second sentence out loud at the top of the call:

   Heads up: I use an AI assistant that transcribes this call locally
   to generate my notes. Nothing is shared outside my team, and I'm
   happy to turn it off — just say the word.

6. Test it against a recording before you use it live. Play a recorded call through your system output and watch what the overlay produces. Tune the prompt until the interrupt rate feels right. Then verify the overlay is genuinely hidden: start a Zoom call with yourself, share your screen, and confirm the window doesn’t appear in the shared view or the cloud recording.

7. Route the output somewhere durable. Post-call summaries that live in an app you never open are worthless. Pipe them to whatever you actually read:

   curl -X POST "$SLACK_WEBHOOK_URL" \
     -H 'Content-Type: application/json' \
     -d '{"text":"*Call: Acme Corp*\n'"$(cat ~/halo/last-summary.md)"'"}'

How it compares: Halo vs Cluely vs the bot notetakers

Capability	Halo / Glass	Cluely	Otter, Fireflies, Fathom	Zoom AI Companion
Joins call as a bot	No	No	Yes	No (native)
Blocked by IT bot policies	No	No	Yes	No
Open source client	Yes	No	No	No
Self-hostable / BYO key	Yes	No	No	No
Reads screen content	Yes	Yes	No	Limited
Real-time in-call answers	Yes	Yes	Minimal	Limited
Hidden from screen share	Yes	Yes	N/A	N/A
Automatic consent disclosure	No	No	Yes (bot is visible)	Yes (platform notice)
Works on any app, not just meetings	Yes	Yes	No	No
CRM / post-call integrations	Early	Early	Mature	Zoom-native

The honest read: bot notetakers still win on post-call workflow — CRM sync, searchable archives, team-wide libraries built over years. Halo and Cluely win on the live layer and on surviving IT policy. They aren’t the same product yet, and many teams will run both until one side closes its gap.

What’s next

Watch the platform response first. Zoom, Microsoft, and Google all have the technical means to make screen-layer capture harder: DRM-style protected video paths, audio-tap detection, meeting-level attestation that no local capture is running. Microsoft already ships protected-content flags that black out captured windows in Teams. If any of them decides invisible desktop assistants are a compliance liability for their enterprise customers, the whole category inherits a much harder engineering problem overnight. My bet is they escalate detection rather than block outright, because too many legitimate tools — screen recorders, accessibility software, support tooling — live in the same API surface.

Second, watch whether the open source client stays meaningfully open. The pattern in this category is predictable: the client goes MIT, the interesting parts — the org context layer, the prompt orchestration, the model routing — stay in a closed backend, and “open source” gradually means “you can read the UI code.” Whether Glass’s capture pipeline and prompt logic remain in the public repo as the team tier matures is the test of whether the transparency claim holds. Check the commit history in six months, not the README today.

Third, the legal layer will move faster than the product layer. Two-party consent statutes were written for tape recorders, and courts have not settled whether an on-device AI assistant that transcribes but doesn’t store audio constitutes recording. The first employment lawsuit involving an invisible AI assistant in a performance review or a termination call will set more of this category’s direction than any feature release. Enterprises will build policy around that ruling, and vendors will build features around that policy.

Frequently Asked Questions

Is Halo actually undetectable in a Zoom or Teams call?

The overlay is excluded from screen-capture APIs, so it doesn’t appear in a shared screen or a cloud recording, and there is no bot in the participant list. That is not the same as undetectable — proctoring software, managed-device endpoint agents, and process monitoring can all see a running application. Treat “invisible to the meeting platform” as the accurate claim and “invisible to your employer’s IT” as false.

How is Halo different from Cluely in practice?

Technically they are close cousins: both are screen-and-audio-layer overlays with real-time answers. The Halo vs Cluely distinction is packaging and trust. Halo ships an open source client you can audit and self-host with your own model keys; Cluely is closed and hosted. If your blocker is security review, that difference is decisive. If your blocker is answer quality on live calls, test both — the gap there is small and moves with every model update.

Do I legally have to tell people I’m using it?

In two-party (all-party) consent jurisdictions, yes, if it records or transcribes — and several US states plus most of the EU fall in that bucket. Even where one-party consent suffices, undisclosed use of an AI meeting copilot without a bot is a trust problem that costs more than it saves when discovered. Disclose by default. The scripted line in step five takes four seconds.

Can I use this as a real-time interview assistant?

Technically yes; that is a large part of why the category got attention. Practically, it’s a bad trade. Employers are adapting — live coding on their hardware, in-person final rounds, follow-up questions designed to expose scripted answers — and getting caught ends the process and often the relationship with that company permanently. The defensible uses are calls where you are accountable for the outcome: your sales calls, your support escalations, your own team’s meetings.

Does the open source version give me everything the paid tier does?

No. The Glass client is open source; the team tier adds shared prompt libraries, org-level context ingestion, admin controls, and hosted infrastructure. Self-hosting the client with your own API key gets you the core real-time assistant and full control of your data. You give up the collaboration layer and take on the ops burden of running it — a reasonable trade for an individual or a small technical team, a harder one for a twenty-person sales org.

What happens to my existing notetaker if I adopt this?

Keep it, at least for now. Halo and its peers are strong at the live layer and thin on post-call workflow — CRM writeback, searchable team archives, integrations built over several years. The realistic 2026 setup for most teams pairs an open source meeting notetaker or overlay for real-time assistance with an incumbent for the record of what happened. Revisit in two or three quarters, when one side has closed the gap.