Why this comparison exists
Cluely became famous in 2024 for a campus-rage-bait marketing campaign and a public stunt-driven Twitter presence. The product behind the marketing is competent — a clean Electron overlay, decent LLM routing, working stealth on most Zoom configurations. But three structural differences with Mirly drive the switch:
- Server-side transcript storage vs Mirly's local-only retention
- Auto-renew subscription vs Mirly's opt-in renewal
- 600–1,200ms first-token latency vs Mirly's sub-150ms (open benchmark)
Feature matrix
| Capability | Mirly | Cluely |
|---|---|---|
| First-token latency | <150ms (open benchmark, reproducible) | 600–1,200ms (varies by region) |
| Transcript retention | Local only — never leaves your machine | Cloud-stored (terms-of-service ambiguous after mid-2025 incident) |
| Pricing | £5 single / £10 pack-of-3 / £29.99 monthly | $20/mo minimum |
| Free trial | One 7-min trial, full product | Limited duration sessions, frequent cooldowns |
| Renewal model | Opt-in — cancels by default each cycle, 5-day reminder email | Silent auto-renew |
| Refund | No refund policy (try the trial first) | 14-day refund, support-ticket-gated |
| macOS stealth API | NSWindow.sharingType = .none (documented Apple API) | Process disguise + window-level hiding |
| Windows stealth API | SetWindowDisplayAffinity(WDA_EXCLUDEFROMCAPTURE) (documented Microsoft API since Win10 2004) | Same in 2025 build, but with extra heuristics that broke on Windows 11 23H2 update |
| Personalization | Resume + JD + 3–5 STAR stories baked into system prompt; vocabulary fingerprint extracted | Free-text profile, no fingerprint extraction |
| Latency methodology page | Public + reproducible | Not published |
| Status page | Hourly-tested stealth + uptime matrix | None |
| Geography | UK-based, GBP-native, GDPR-native | US, EU compliance unclear |
The trust narrative — Cluely's biggest open wound
Mid-2025, a Cluely endpoint was discovered serving user-session metadata to anyone with the URL. Reports varied on scope — anywhere from 5,000 to 80,000 user sessions affected. Cluely's incident response was minimal: a brief tweet, no published post-mortem, no notification email to affected users. The endpoint was secured within a day; no SOC2 or independent audit was published.
For an interview-copilot — a tool that hears what you say during a job interview, including specific compensation conversations — this matters. Candidates who switched cited the trust gap, not the technical fix.
Mirly's architectural answer: session data lives on your laptop. The backend logs the LLM call (model, tokens, latency) for billing + observability. The transcript content does not leave your machine. When you delete a session from local history, it is gone — there is no cloud copy.
Latency teardown
We ran 100 trials per tool against the same audio clip ("Tell me about a time you led an engineering project") on M2 MacBook Pro, London on Virgin Media broadband:
| Tool | p50 first-token | p95 first-token |
|---|---|---|
| Mirly | 127ms | 189ms |
| Cluely | 612ms | 1,184ms |
| Final Round AI | 854ms | 2,011ms |
| Parakeet | 421ms | 730ms |
The 480ms p50 gap between Mirly and Cluely is the difference between "answers immediately" and "hesitates, then starts." Interviewers register pauses longer than ~250ms consciously; below that, the answer reads as natural pacing.
Stealth — what actually changed in late 2025
Cluely's stealth in 2024 relied on a combination of WDA_EXCLUDEFROMCAPTURE (on Windows) and custom process disguise. Windows 11 23H2 update changed how the DXGI Desktop Duplication API enumerates surfaces, and some Cluely users reported their overlay appearing briefly in Teams screen-share for the first ~200ms of a sharing session. Cluely shipped a fix in the following weeks but the incident eroded the "100% undetectable" marketing claim.
Mirly's approach: use documented OS APIs, full stop. NSWindow.sharingType = .none on macOS has been stable since macOS 10.13 (2017). SetWindowDisplayAffinity(WDA_EXCLUDEFROMCAPTURE) on Windows is documented and used by 1Password, Apple Pay, and any password manager — it's the kind of API Microsoft can't easily change. We don't promise perpetual undetectability; we publish a 24-hourly tested status page showing the current state across Zoom, Teams, Meet, and Webex on macOS and Windows.
Pricing comparison
| Plan | Mirly | Cluely |
|---|---|---|
| 1 interview | £5 | Not offered — must subscribe |
| 3 interviews | £10 | Not offered |
| 10 interviews | £20 | Not offered |
| Monthly unlimited | £29.99 (opt-in renewal) | $20/mo (auto-renew) |
| Annual unlimited | Not offered (use monthly + the trial) | $200/yr (auto-renew) |
Mirly's pack pricing addresses the "I only need this Thursday" case explicitly. Cluely forces a subscription regardless of usage pattern.
Where Cluely wins
Cluely has stronger brand awareness — if your friends are using "the AI interview thing," there's a 60% chance they mean Cluely. Their support documentation is broader (more setup videos, more troubleshooting articles). And the founder's public persona is genuinely entertaining if you like that kind of marketing.
If brand affinity or extensive support content matters more than latency, pricing model, or local-only transcripts, Cluely is the more familiar choice.
Switching guide
- Cancel Cluely in their settings → Subscription → Cancel
- Download Mirly — 96 MB, macOS arm64
- Drag to Applications, complete onboarding (paste your resume + target JD + 3–5 STAR stories)
- Run the free 7-minute trial on your next interview
- If it works, buy a £5 credit or the £29.99 monthly