Your restaurant.
In plain English.
Tammy doesn't open a dashboard. She types into Claude Desktop on her laptop — “how's tonight?” — and gets a real answer from her POS, scoped to her restaurant only. 19 tools. 60-second onboarding. Bidirectional knowledge moat.
›
19
Tools live in production
< 60s
Tenant onboarding time
15+
Cross-restaurant patterns
What she types
Real questions, real tools, real answers
Every question routes to the right tool, scoped to her restaurant. Tenant keys never see another restaurant's data.
› “How's tonight at Chan Nara?”
$2,847 on 38 orders so far. Avg ticket $74.92. Top: Pad Thai (12). Up 18% vs same time last Tuesday.
→ air_today_summary
› “How's tonight vs same day last week?”
Tonight is up: $2,847 vs $1,920 (+48.3%). 38 orders vs 24 (+58%). Top item swapped: Pad Thai displaced Tom Yum.
→ air_compare_periods
› “Who hasn't been in for a month?”
12 lapsed regulars, ranked by lifetime spend. Top: Sarah K. — $847 lifetime, last visit 34 days ago. 8 reachable by SMS.
→ air_who_came_in
› “Why is my prime cost up?”
70% for 3 weeks puts you in distress territory. Pull food cost % and labor cost % separately. Per the Prime Cost Playbook, the most common culprits are labor bloat or portion drift…
→ air_ask_cfo
› “86 the salmon”
Salmon Sashimi is 86'd. Disappears from POS, website, and online ordering immediately.
→ air_eighty_six_item
› “Set the omakase to $48”
Omakase: $42.00 → $48.00. Square sync picks this up within ~60 seconds.
→ air_set_price
› “Make pad thai today's special, 12 portions”
Pad Thai is now today's special with 12 portions. Limited-quantity specials auto-86 when remaining hits 0.
→ air_set_special
› “Drop these 50 photos and match each to my menu”
Vision-matched 47/50 to existing items. Salmon Sashimi attached. Pad Thai attached. (3 ambiguous — please disambiguate.)
→ air_upload_image
› “Text my regulars about tonight's pad thai special”
Preview: 87 opted-in members in audience. Confirm to send. Reply STOP auto-handled.
→ air_send_announcement
› “How do I troubleshoot the printer?”
Loaded restaurant/onboarding-troubleshooting-runbook. Step 1: check Wi-Fi. Step 2: try a different iPad. Citing 12,863-byte runbook from prior go-lives.
→ air_search_patterns + air_get_pattern
› “Save this fix as a pattern”
Saved as [PROPOSED] in the AI-R team's review queue. Next restaurant onboarding will see this once promoted.
→ air_propose_pattern
› “What's been happening on AIR today?”
37 tool calls, 0 failures. Top: air_today_summary ×12. Tammy 86'd the salmon at 3:42pm.
→ air_get_activity
How it works
Two systems. One conversation. Tenant data never escapes.
Operators live in AIR. Orion is the institutional brain. The seam is a per-tenant scoped MCP key + narrow service keys for cross-system reads.
Tenant surface (AIR)
Operators live here
- • Claude Desktop / Cowork
- • 19 AIR-MCP tools
- • Per-restaurant scoped key
- • Audit log per call
- • Self-serve subscriptions
Server-to-server bridge
narrow-scope service keys
Institutional brain (Orion)
AI-R team + IP moat
- • 15+ patterns with gotchas
- • Onboarding runbooks
- • 12 CFO playbooks (~237 RAG chunks)
- • Decision log + narratives
- • Cross-restaurant benchmarks
Bidirectional knowledge moat
Patterns + runbooks accumulated across every restaurant the platform has ever served. Tenants consume them via air_search_patterns. Tenants contribute back via air_propose_pattern. Reviewed + promoted by the AI-R team. Compounds with every onboard.
15+
patterns + runbooks live
Cross-system intelligence bridge
When a tenant asks 'why is my prime cost up?', AIR's server proxies into Orion's CFO playbook agent (12 long-form playbooks, ~237 RAG chunks). Tenant gets the answer with citations. Tenant key never reaches Orion.
$0.017
per CFO answer with citations
60-second onboarding
Spin up the next restaurant via one CLI command. Provisions the restaurant row, loyalty program with default rewards, owner role, AIR-MCP key, draft cutover checklist on Orion.
< 60s
from CLI to ready-for-Claude-Desktop
Audit + TCPA-grade compliance
Every tool call logged with caller, restaurant, args preview, success/failure, duration. SMS sends require explicit opt-in; preview/confirm pattern + 5-min token expiry; STOP-webhook auto-applies opt-outs from Twilio inbound.
100%
of tool calls audit-logged
Ready to give your operators this?
We're onboarding our next 4 restaurants in May. Each gets a private operator key for Claude Desktop / Cowork on day one.
For investors: full architecture deep-dive available on request — neil@ai-restaurant.net