The deliberate alternative to OS-bundled AI.
Microsoft, Apple, Google and NVIDIA are racing to bake AI into every OS and device — bundled, all-or-nothing, vendor-locked. Concierge is the opposite: opt in to only the modules you need, pick whichever model works best today, keep your accounts and your data where you want them — on your hardware, or on ours.
Concierge is what happens when you build a personal AI assistant around the user's needs, not the vendor's. One app pulls together email, calendar, meetings, research, and creative work — using whichever model is best for the job today.
The timing matters: Microsoft Copilot+ PCs are already shipping. Apple Intelligence is live. Google Gemini Nano is in Android. NVIDIA's DGX Spark workstations are bringing 200B-param local models to desktops. Within 18–24 months, every consumer device will have an OS-bundled AI. That captures the casual mainstream — but every one of those products exists to lock the user deeper into a single vendor's grip. The harder Big Tech pushes integrated AI, the larger the audience grows for the deliberate alternative.
Microsoft Copilot is glued to Office. Apple Intelligence assumes iCloud. Google's deep in Gmail. None will integrate cleanly with the competitor's services — that's the business model, not a bug.
OS agents ship as a single product. You can't turn off Recall but keep search, or use Claude for reasoning and Apple Vision for images. The bundle IS the strategy.
Today's best model is tomorrow's also-ran. OS-integrated AI ties you to whatever the vendor chose when they shipped the device.
Power users have iPhone, Windows laptop and Linux server. OS agents do not span all three coherently — and never will.
Same structural shape that made Signal viable next to iMessage, Brave next to Chrome, Proton next to Gmail. The audience that actively rejects vendor lock-in grows as the giants get more aggressive — and they will pay for the alternative.
A concrete picture of what a power-user's day looks like with one assistant orchestrating their stack — instead of jumping between five tools and remembering what's where.
"What's my day look like?" — Concierge pulls today's calendar, the inbox since you logged off, last week's notes, and overdue follow-ups. Drafts you a focused written brief: priorities, conflicts to resolve, what to tackle first.
It listens to your Meet/Zoom call, transcribes the discussion with speaker diarization, pulls action items into your task system with owners and due dates. No more screenshotting notes afterwards.
47 new emails since lunch. Concierge sorts them — invoices to forward, replies it can draft in your voice, newsletters to skip, urgent threads to surface. You skim a curated list, approve drafts in bulk.
"Generate three thumbnail options for the blog post — moody, technical, friendly." Nano Banana wired straight into the chat. Pick one, drop it into your CMS without leaving the conversation.
"What changed in EU AI regulation this month?" Concierge runs an autonomous multi-step research pass — gathers sources, reads them, synthesises a written brief with citations. Thirty minutes later you have a report, not 50 browser tabs.
One install. One app. Modular by design: toggle on the integrations you want, leave the rest off. Backed by whichever model gives you the best price-to-performance — Claude, GPT, Gemini, or a local Llama for offline.
IMAP/SMTP. AI summaries, style-matched replies, auto-tagging.
CalDAV. Smart scheduling, meeting prep, conflict detection.
Multi-model. Memory across sessions. Voice in/out.
Tool use, web browsing, file ops. "Plan my week" not just "answer me."
Whisper + diarization. Summaries, action items, follow-ups.
Nano Banana / Flux / SDXL APIs. In-line with chat.
Multi-step gathering + synthesis into a written brief.
Veo 3 / Sora APIs. Queue-aware. Hold for cost/latency to settle.
| Concierge | OS-integrated Copilot+ · Apple Intel. · Gemini Nano |
Cloud AI assistants ChatGPT · Claude.ai · Perplexity |
Productivity-suite AI M365 Copilot · Workspace AI |
Self-hosted chat Open WebUI · LibreChat |
|
|---|---|---|---|---|---|
| Modular opt-in | ✅ Pick what you need | ❌ Bundled by vendor | ❌ Single product | ❌ All-or-nothing | ❌ Chat only |
| Self-host OR managed | ✅ Both, customer choice | ❌ Vendor-controlled | ❌ Vendor-hosted only | ❌ Vendor-hosted only | ⚠️ Self-host only |
| BYO accounts (cal / email) |
✅ Your accounts, we orchestrate | ❌ Locked to vendor ecosystem | ⚠️ Plugin-limited | ❌ Tied to suite (Outlook / Gmail) | ❌ Not integrated |
| BYO model | ✅ Claude, GPT, local — swap freely | ❌ Vendor-chosen | ❌ Vendor-locked | ❌ Vendor-chosen | ✅ Yes, chat only |
| Cross-platform | ✅ Same product everywhere | ❌ Per-OS silos | ✅ Web / iOS / Android | ⚠️ Within suite ecosystem | ⚠️ Web-based only |
| App-shaped UX | ✅ Built for non-developers | ✅ Polished | ✅ Polished | ✅ Polished | ❌ Developer-feeling |
| Source transparency (strategic choice — see below) |
⚙️ Open-core / source-available / proprietary — TBD | ❌ Closed | ❌ Closed | ❌ Closed | ✅ Open |
| Multi-modal generation (image / video / voice) |
✅ Any provider — Nano Banana, Veo, ElevenLabs | ⚠️ Vendor-chosen, often paywalled | ⚠️ Vendor's stack only (DALL-E, etc.) | ⚠️ Limited & vendor-controlled | ❌ Chat only |
| Tool extensibility (MCP / custom skills) |
✅ MCP-native, add any tool | ❌ Vendor-curated only | ⚠️ Plugins / Actions, curated | ⚠️ Within suite only | ⚠️ Partial |
| Cost transparency | ✅ BYO API key — see every cent | ❌ Bundled into device / sub | ⚠️ Subscription, surprise API costs | ❌ ~$30/user/month, opaque | ✅ Self-host costs only |
| Team / household sharing | ✅ One instance, multi-user | ❌ Per-device / per-account | ⚠️ Per-seat team plans | ⚠️ Per-seat, vendor identity | ⚠️ Possible but rough |
We run it. They open the app and use it. Standard SaaS model.
One-shot installer. Their hardware, their network, their data.
Deployment (self-host vs hosted) is independent of source-availability. The licence model is an open question worth deciding early.
Open-source projects already cover ~60% of the technical surface. The remaining 40% — the differentiated product — is the polish, modularity, deployment story, and integration depth.
PewDiePie's MIT-licensed self-hosted AI workspace (32k+ GitHub stars in days, May 2026). FastAPI + JS. Agent loop, MCP support, hardware-aware model serving. Realistic starting point or reference.
The full Claude Code system prompts, tool descriptions, and subagent dispatch patterns are publicly documented in maintained GitHub repos. Years of Anthropic's UX research, available to learn from.
Whisper for transcription, CalDAV/IMAP libraries, Stripe for billing, OAuth flows for each provider — all mature, well-supported, no novel research required.
Stand up Odysseus on a test box. Wire one module deeply (email triage). Confirm the agent loop hits our quality bar with Claude as backend. Output: a working prototype + a sharper market read.
Polished onboarding (OAuth wizard for calendar / email / model). 4 modules live: chat, email, calendar, autonomous agent. PWA shell. One-line installer for the self-host path. Internal alpha across the team.
Friends-and-family hosted users plus a handful of self-host community testers. Iterate hard on onboarding. Add meeting transcription + image generation as Fast Follow modules.
Execute the chosen commercial model. Launch the managed tier. Distribution: HN, ProductHunt, the natural privacy-first communities. Measure conversion, iterate on pricing tiers.
Their roadmap is public; we monitor and stay ahead on the modules + deployment story. We're not racing on chat features (they win those), we're racing on personal-assistant polish.
Hard per-user quota limits from day one. Hybrid model: free tier uses local cheap model, paid tiers add Claude/GPT quota. Tier ceilings prevent surprise bills.
Start with 2 providers per module (Google + IMAP for email; Google + CalDAV for cal). Expand only when a tier-1 customer asks. Resist feature-creep.
Real but slow. Our privacy + self-host story is the moat they cannot match. Our customers are the ones actively opting out of those ecosystems.
Position as productivity tool first, AI second. "Replace 5 SaaS subscriptions with one assistant" is a clearer pitch than "look, more AI."
If the underlying bet holds — that there's a real audience for the deliberate alternative to OS-bundled AI — these are the threads worth pulling.
A short feasibility spike — a working prototype with one module wired deeply — would answer most of the technical questions cheaply and produce something real to react to.