Volume I · Chapter 01 · Mercer Street, 2026mshdash.com
Nobody readsevery ticket.Until now.
MSH Intelligence captures every line from every POS across every venue — Omnivore, Salido, Micros, Shift4 — joined to every booking from SevenRooms, and turns it into plain-English briefings, conversational forensics, and a raw archive your POS does not keep. One mind, reading the whole house.
Bowery Meat Company East Village · NYC·Lure Fishbar NY SoHo · NYC·ChaChaTang Chinatown · NYC·Bar Mercer SoHo · NYC·Smyth Tavern TriBeCa · NYC·Galerie SoHo · NYC·Lure Fishbar Chicago River North · CHI·Seahorse Seaport · NYC·Bowery Meat Company East Village · NYC·Lure Fishbar NY SoHo · NYC·ChaChaTang Chinatown · NYC·Bar Mercer SoHo · NYC·Smyth Tavern TriBeCa · NYC·Galerie SoHo · NYC·Lure Fishbar Chicago River North · CHI·Seahorse Seaport · NYC·
§ 01 · the problem
Your numbers are already here. You just can’t see them.
I · The blindness
Eight rooms, five silos.
Omnivore. Salido. Micros. Shift4. SevenRooms above them all. Each vendor exports its own shape of data on its own schedule. No shared schema. No shared calendar. No single place where yesterday actually lives — and no place where the booking ever meets the check.
II · The delay
By the time the report arrives, the shift is cold.
Patterns repeat while you sleep. Leaks compound for a full week before anyone notices. Managers report by memory; memory forgets the ticket that mattered.
III · The ask
Somebody should read every comp. Nobody has the hours.
Every void, every house discount, every Event-Menu line, every covers-to-revenue mismatch. That’s the GM’s job. That’s also the GM’s third job. So it waits.
§ 02 · the layer
One intelligence, sitting above every POS.
01
It ingests every POS — and the reservation layer that sits above them.
Omnivore REST every two hours, odd UTC. Salido CSVs nightly via SFTP. Micros direct where we can get in. Shift4 outreach in flight. SevenRooms reservations every two hours across all eight venues, joined by check_number to the underlying POS ticket. The layer doesn’t care what vendor you buy from.
02
It captures the raw ticket — not the rollup.
Every closed check, every line item, every employee id, every revenue center — upserted into raw_tickets / raw_checks / raw_check_items with the full JSONB payload. Forever. Because Omnivore purges, and Salido exports end.
03
It reads the night at 4 a.m. and hands you the brief at 6.
Claude Haiku 4.5 writes a one-page briefing after every sync: what jumped, what dropped, what to ask about at the stand-up. Plain English. Dated. Archived.
04
It answers questions in conversation — with the receipts.
Claude Sonnet 4.6 with tool-calling against the warehouse. Ask in English. Get back numbers with citations to the exact tickets they came from. Session memory scoped per user.
05
It lives inside any AI assistant, via MCP.
mshdash.com/api/mcp exposes the full warehouse through the Model Context Protocol, OAuth 2.1 secured. The same data, answerable from Claude Desktop, ChatGPT, a phone, or a scheduled agent.
§ 03 · architecture
The layer, diagrammed honestly.
Upstream · the house
Bowery Meat · ChaChaTang · Lure NYOMNIVORE API
Bar Mercer · Smyth · GalerieSALIDO · SFTP
Lure ChicagoMICROS RES 3700
SeahorseSHIFT4
All eight venues — bookings, covers, no-shows, VIPsSEVENROOMS · API
Claude Haiku 4.5 — written at 03:20 EST, delivered at 06:00.
Conversational
Claude Sonnet 4.6 — tool-calling against the warehouse.
Dashboard
mshdash.com — portfolio, location, menu, exports.
MCP
/api/mcp — the warehouse, as an MCP server.
§ 04 · capabilities
Six chapters of what it does for you, every day and every night.
Each of the following is in production and running over MSH data today. The copy and callouts below describe capability — the numbers in the demo cards are illustrative.
I.
Portfolio
Eight rooms, one pane.
Yesterday rolled up, broken down, and placed beside the same night last week. Revenue, covers, average check, mix — joined to bookings from SevenRooms so reserved covers, walk-ins, no-shows, and VIP visits land in the same view as the tickets they produced.
portfolio viewlocation comparecovers vs actuals7-day trendcsv export
II.
Nightly briefing
A one-page brief, on your desk by six.
Claude Haiku 4.5 reads the night the moment it syncs. What jumped. What dropped. What to ask about at the morning stand-up. Written in plain English, dated, signed.
generated 03:20 ESThaiku 4.5 via ai gatewayarchived nightly
III.
Conversational forensics
Ask the question.
Claude Sonnet 4.6 answers in plain English, with the tool-calls to back it up. “Which server had the highest comp rate at Bowery last week?” “Which menu items are underperforming their category average at Galerie?” Full access to the warehouse. No SQL required.
sonnet 4.6tool-callingper-user session memory
IV.
Raw archive
The tickets the POS forgets.
Every closed ticket, every line, every employee, every revenue center — captured nightly and kept forever. Omnivore purges on a rolling window. Salido exports end. The archive doesn’t. This is the memory the operators never had.
The forensics a GM would run — if a GM had the hours.
Comp totals by employee. Void totals by shift. Event-Menu nights flagged automatically. No-show rates by booking channel. VIP arrivals attributed to the server who took them. Anomalies surfaced before the next prep sheet is printed.
mshdash.com/api/mcp exposes the warehouse as a Model Context Protocol server — eighteen tools, five power prompts, including SevenRooms guest history and reservation querying. Connect it from Claude Desktop or claude.ai on your Max plan, and Opus 4.7 with 1M context runs the deep forensics at zero marginal cost — the flagship model you already pay for, now pointed at your numbers. (Full breakdown below.)
18 tools5 promptsOpus 4.7 via Maxno API bill
§ 05 · the mcp layer
Bring your own Opus.
The dashboard’s in-app Ask runs on Claude Sonnet 4.6 — the right call for an always-on feature that bills per token. But when the question is big — find every revenue leak across every venue for the last ninety days — you want the flagship model doing the work. Connect the MCP to Claude Desktop or claude.ai on your existing Max plan and Claude Opus 4.7 with a 1M-token context window takes the question. Same data. Deeper reasoning. The conversation runs on your Claude subscription — so the heavier investigations don’t add to the dashboard’s AI bill.
In-app Ask · Dashboard
Claude Sonnet 4.6
Fast, always-on, shared across users
Paid by the operator, per token
200k-token context window
Right tool for everyday questions
Via MCP · Your assistant
Claude Opus 4.7
Anthropic’s flagship model — deepest reasoning available
Runs on your Max / Team subscription — doesn’t add to the dashboard’s AI bill
1M-token context window — an entire quarter in one prompt
Right tool for the investigation that actually matters
mshdash.com/api/mcp is a production Model Context Protocol server. Plug it into any MCP-speaking assistant — Claude Desktop, claude.ai, ChatGPT with connectors, Cursor — and the full warehouse becomes queryable in plain conversation. Ask from the office. Ask from the train. Ask the AI in your pocket to run last night’s forensics while you’re still on the sidewalk.
Paste into Claude Desktop → Settings → Connectors → Add custom MCP server. Consent flows through your existing MSH dashboard login — no new credentials to provision, no API keys to rotate. Access revokes the instant a user account is disabled.
Ask.
things you can say, in plain english
“How did we do yesterday across all the venues?”
“Compare this Saturday to the same Saturday last month at Bowery.”
“Which server had the highest comp rate at Galerie last week?”
“What are my slowest nights? Give me three ideas to fix them.”
“Show me a chart of revenue by venue for the last 90 days.”
“Read me the briefing from last Tuesday.”
“How are wines-by-the-glass moving at Galerie this quarter?”
“Show our no-show rate at Lure NY by booking channel for the last 30 days.”
“Pull the guest history for the 7pm reservation at Bowery tonight.”
“Compare booked covers to actual covers across all venues this week.”
“Draft a manager note based on what you just found.”
No SQL. No data-team ticket. Ask the way you’d ask a sharp colleague who’s already read the numbers.
One tap.
five investigations, preloaded
/ morning briefingA one-page brief for any day in the archive — same voice as the 6 a.m. email, on demand.
/ compare weekHow is this week tracking against last week, across every venue that has enough data to compare honestly?
/ find revenue leaksThe forensic sweep — comps, voids, suspicious nights, servers running hot. Every finding cites the exact tickets behind it.
/ find opportunitiesWhere could you be making more? Slow-day activation, buyout systematization, menu engineering, cross-venue benchmarks.
/ openA guided tour of everything this thing can do, with example questions you can copy and try.
Think of these as pre-loaded investigations. Tap one, get a multi-step workup — no prompt engineering required.
Power.
the upgrade path for the questions that matter
Runs on Claude Opus 4.7Anthropic’s most capable model — deeper reasoning, bigger memory, better at reading the patterns that matter.
Holds a whole quarter in its headA 1-million-word context window means you can hand it three months of data in one conversation and it will remember all of it.
Uses your own Claude subscriptionWhen you chat via Claude Desktop or claude.ai, your Max or Team plan runs the conversation. Heavier investigations don’t lean on the dashboard’s AI budget.
Works everywhere you chatClaude Desktop on your Mac. claude.ai in a browser. ChatGPT with custom connectors. Even on your phone.
Signs in with MSHNo new password. No API key to paste. Your existing MSH dashboard login grants access — and revokes it the instant your account is disabled.
Think of the dashboard as the always-on intelligence — briefings, the in-app Ask, the sync pipeline. The MCP is the flagship upgrade you reach for when a question really warrants Opus.
A typical conversation · illustrative
F train · Tuesday · 8:14 a.m. — a GM chats with the AI on their phone:
How are the three NYC places trending? Give me a chart for the last 90 days.
Pulling ninety days for Bowery, Lure, and ChaChaTang. One moment.
Reading revenue by day, three venues…
Building a chart for you…
Revenue · last 90 days · NYC portfolioillustrative
Bowery Meat $44.8k Lure Fishbar NY $30.6k ChaChaTang $25.9k
Bowery’s trend is the strongest of the three — a clear upward walk over the last six weeks, with three standout Fridays. Lure is steady but flat. ChaChaTang has been climbing slowly since week five. Want me to break any of these down by day-part, or pull the underlying numbers as a spreadsheet?
Break Bowery down. What made those Fridays different?
Looking at the three standout Fridays at Bowery.
Reading the ticket-level detail for those three nights…
Checking for private events, weather, menu features…
Two of the three were private buyouts — a single large check each, booked through the events calendar. Strip those out and the dining-room revenue was in line with an average Friday. The third Friday was a regular service with a heavier-than-usual bar mix — whiskey flights and champagne-by-the-glass ran roughly fifty percent above their trailing average. Possibly a walk-in group; worth asking the bar manager.
Want a manager note drafted for tomorrow’s stand-up that summarizes this?
Yes, and email me the weekly numbers behind it.
Drafting the note now — I’ll keep it short enough to read at a stand-up, and I’m putting together a spreadsheet with the weekly revenue breakdown for your inbox.
Writing the manager note…
Preparing the spreadsheet…
Powered by Claude Opus 4.7 on the GM’s Max plan · no dashboard open · no laptop required
§ 06 · a taste of the output
What the morning looks like.
Yesterday · Portfolio · A sample morning
Revenue
$68,412
+6.4% vs. same day last week
Covers
452
+2.1% vs. same day last week
Avg check
$151.35
-2.6% vs. same day last week
Bowery Meat led the portfolio on the strength of an Event-Menu buyout; strip that out and the dining room tracked in line with the trailing four Wednesdays. Galerie covers ran ahead of revenue — average check compressed roughly 6%, consistent with a weather-driven shift toward apps and wines-by-the-glass. ChaChaTang comp rate ticked above the trailing-seven average; one server accounted for most of the variance and should be reviewed.
— Claude Haiku 4.5 · 03:21 EST · brief-2026-04-15
An example conversation.
Which server had the highest comp rate at Bowery last week, and which checks drove it?
Looking at last week’s checks at Bowery…
Comp rate at Bowery Meat last week was meaningfully led by [server A], roughly two-and-a-half times the house average across six shifts. Three checks above eight-percent comp accounted for most of it — all large parties paid partially by house promo. I’ve got the check numbers ready if you want them.
Do the same for ChaChaTang for the last thirty days and send me a spreadsheet.
Running the same look across ChaChaTang…
Putting a spreadsheet together…
On it — ranked list coming and the spreadsheet is on its way to your inbox.
§ 07 · the house
The eight rooms of Mercer Street.
01
Bowery Meat Company
East Village · NYC
Omnivore
● in the layer
02
Lure Fishbar NY
SoHo · NYC
Omnivore
● in the layer
03
ChaChaTang
Chinatown · NYC
Omnivore
● in the layer
04
Bar Mercer
SoHo · NYC
Salido
● in the layer
05
Smyth Tavern
TriBeCa · NYC
Salido
● in the layer
06
Galerie
SoHo · NYC
Salido
● in the layer
07
Lure Fishbar Chicago
River North · CHI
Micros RES 3700
○ inbound
08
Seahorse
Seaport · NYC
Shift4
○ inbound
Six venues live for POS data today — Bowery Meat, Lure NY, ChaChaTang, Bar Mercer, Smyth Tavern, Galerie. SevenRooms reservations flow for all eight, including Lure Chicago and Seahorse while their POS integrations come online. The remaining POS work is already mapped.
§ 08 · what this is, what it is not
This is —
A forensic operational layer.
A nightly briefing on your desk by six.
An always-on analyst for every venue.
A raw archive your POS refuses to keep.
A system that answers, in plain English.
This is not —
Another dashboard nobody opens.
Another PDF emailed too late.
Another login for the GM.
Another tool waiting for its data.
Another generic AI wrapper.
Return to the workyou actually love.
Leave the tickets, the comps, the voids, the reconciliations, the nightly math to the machine. Read the one page it writes you at six.