Twenty-seven tools, read for a build. This opens with the analysis — where the market is, who leads each job, and the practices worth copying — then the full filterable catalogue sits in Appendix A, each tool as a faithful mockup of its layout family (not a screenshot) with a link to its site.
"Name-card scanner slash work CRM" sits across three categories people routinely conflate. The strongest builds treat them as separable layers.
Paper card → structured contact. OCR & the review screen are everything.
Win condition: parse accuracy + correction speed
Keep people warm: history, reminders, "you haven't spoken in 3 months."
Win condition: you actually open it → low-friction capture
Scan → enrich → qualify → sync to CRM. Built for sales & events.
Win condition: lead-to-CRM latency (hours not days)
The strategic read across all 27 — and what it implies for a custom build.
Capture has been commoditised. Apple Live Text and Google Lens already parse a card for free, on-device, offline. A tool that only scans has no moat — so across the category the value has moved downstream: qualify, dedup, nudge, and sync. Apps that stop at raw capture leave you a list you still have to clean by hand.
The category is littered with corpses. ABBYY killed its cloud sync in 2023; BizConnect stopped shipping. "Best scanner" listicles still recommend both. Active development is itself a selection criterion — and a reminder that a solo-built tool you control sidesteps abandonment risk entirely.
It splits on three axes. Auto-sync vs manual entry (Dex/Clay/Cloze vs Monica/Notion); organised by person vs project vs pipeline (Dex vs Storyflow vs folk); and cloud vs local-first (Clay vs Monica/Dextr). Where you land on each axis is the whole design.
Paper is declining, not dead — and least dead where you work. Digital exchange (QR/NFC) is rising in tech/SaaS, but manufacturing, healthcare, and most of APAC still run on paper. For institutional finance in Singapore, bilingual EN/CJK cards remain routine — so multi-language OCR and on-device handling matter more for you than for a US SaaS user, not less.
The white space (inference). No incumbent combines on-device / local-first privacy + vision-LLM parsing + a Notion-native canonical store + relationship nudges tuned to a work counterparty network. Dextr is closest on local-first but is Apple-only and not Notion-backed; Clay leads enrichment but is cloud and access-hungry; Dex leads nudges but is LinkedIn-bound. That gap is a credible build thesis — and, given a publish-to-demonstrate strategy, the artifact itself has value beyond the tool.
Build vs buy, honestly. For most people: buy — Dex, Clay, or free Wave Connect cover it. Building is justified here by your stack and the demonstrability of the artifact, provided you scope to the gap: lean on free on-device OCR (Apple Vision) plus a small local or sub-cent model for structuring, and spend your effort on the canonical store, the nudge engine, and the work-context layer no one else builds.
The standout for each job, with its closest runner-up. Mockups show each leader's layout family; names link to the tool.
Distilled from how the category leaders actually behave. These are the patterns worth porting into a build.
If you were buying rather than building — or choosing a reference to study — start here.
The decisions that matter most for a custom build on a Notion / MCP / Apple-hardware stack.
Parsing is two steps, and the cheap split wins. Step one is OCR (image → text); step two is structuring (text → {name, title, company, emails, phones}). Modern phones do step one on-device for free (Apple Vision, Google ML Kit). Step two is then a text task — a few hundred tokens, no image — so it runs on a tiny local model or a sub-cent cloud call. A full vision LLM (sending the image) only earns its keep on hard cards — foil, vertical, dense bilingual — where OCR garbles reading order. Use it as a fallback, not the default.
| Tier | Approach | Per card | Private / offline | Use when |
|---|---|---|---|---|
| 0 · On-device | Apple Vision / ML Kit / PaddleOCR for OCR → rules or a tiny local text model to structure | $0 | Yes / Yes | Default — esp. work contacts |
| 1 · Local VLM | Qwen3-VL 3–7B, PaddleOCR-VL 0.9B or SmolVLM2 on your Mac/desktop via Ollama / LM Studio / MLX, served over Tailscale | $0 (power) | Yes / on-LAN | Hard cards; one private engine for all devices |
| 2 · Cheap cloud | Gemini 3.1 Flash-Lite / Flash, GPT-nano, Claude Haiku — image or text | ~$0.0002–0.002 | No / No | Fallback, or mobile away from your mesh |
The "expensive" worry is mostly a mirage at personal scale. Even Tier 2 with a cheap multimodal model is a fraction of a cent per card — a few hundred cards a year is pennies. Real cost only appears at high volume or if you reach for a frontier vision model (Opus / GPT-5 vision) you don't need. For a private work book, Tier 0/1 is both cheaper and more appropriate.
| Model | Size / runs on | Why for cards |
|---|---|---|
| PaddleOCR · PaddleOCR-VL | 0.9B · runs nearly anywhere | Dedicated doc OCR, strong CJK — fits bilingual Asia cards |
| Qwen3-VL (or 2.5-VL) | 3–7B · Mac Mini (MLX/LM Studio) or HomeAMD GPU | Best small VLM for documents; ~95 DocVQA, 32-language OCR, reads tilted/low-light |
| SmolVLM2 / Gemma 3 | 2–4B · 4GB-class GPU, iPad-feasible | Tiny; fine for structuring + clean-card OCR |
| Apple Vision + on-device LLM | built into iOS / macOS | Free OCR; pair with a small local text model to structure — zero marginal cost, fully offline |
| On-device (Vision / ML Kit / local) | Cloud (server / hosted LLM) | |
|---|---|---|
| Latency | ~100–500ms / card; 50 ≈ 5–25s | ~1–5s / card; 50 ≈ 50–250s |
| Offline | Works — airplane mode, basements | Queues or fails without network |
| Privacy | Data never leaves device | Temp on vendor infra; GDPR/PIPL exposure |
| Accuracy | Strong on clean cards; weaker on foil/dark/vertical | Historically better field mapping |
| Cost | No per-page compute | Per-page cost → scan quotas |
The asymmetry trap: cloud apps let you capture offline but silently queue the parse. The UI looks fine while cards pile up unsynced. On-device parsing avoids this — surface queue state explicitly if you go hybrid.
Strip the branding away and every tool collapses into one of these layout families. Each card in the appendix uses the matching paradigm.