Every return, exchange, and refund in company history β in plain English. Loop is where all our returns and RMAs live. The Loop Returns connector gives everyone instant read access to that world, and gives teammates with Loop credentials the full read/write toolkit β search, analyze, and act on returns without opening a dashboard.
Loop is the returns & exchange platform attached to our Shopify
store β when a customer returns a piece or swaps a size, it happens in Loop (exchanges then land in ShipHero as
EXC- orders). The Loop Returns connector plugs that world into Claude: ask real questions, get real
answers, and β with the right credentials β handle the return end-to-end, all in chat. ("MCP" = Model Context
Protocol, the standard way Claude plugs into outside tools. A "connector" exposes safe, named tools Claude can call.)
Every return and exchange ever: state & outcome, refund / credit / exchange dollars, the customer & original Shopify order, line items with return reasons, labels & live tracking, internal notes, timelines, fraud-risk flags, policies, and the inbound-package report. Real scale: 1,320 returns worth $184,283 in Jul 1β23 alone.
For everyone: search, drill in, and analyze β return rates, reasons, refund $, trends. For Loop-credentialed users: also act β note, flag, approve, reject, process, report fraud, manage the block/allowlist β with Claude showing you exactly what it will do before it does it.
It doesn't live alone. Loop answers join Shopify, shiphero26, Gorgias, Klaviyo, ShopMy, QuickBooks and Asana in a single ask β "which styles our creators push get returned most, and are the exchanges actually shipping?" is one question now, not four dashboards.
The Loop Returns connector shows up as two entries in your connector list. (One carries the analytics engine, the other the search + action toolkit; Claude treats them as one and routes automatically.)
ok: true β then something real: "list this week's open returns."Everyone connects the same way (section 2). What Claude can do for you depends on whether you carry Loop credentials β and, if you do, which role David has assigned you in Loop Admin. Claude can never exceed your own access.
Every MM teammate Β· MM Google account Β· look, search, analyze β can't change anything
CX leads & ops Β· Loop login + role Β· everything at left, plus action
manage-returns. Whatever you could do in the Loop dashboard yourself β
that, and only that, is what Claude can do for you.You never type tool names β Claude routes plain English. Each prompt below exercises one specific tool, so this doubles as the full capability checklist. READ anyone Β· WRITE Loop-credentialed, previewed first
loop_health"Run a health check on the Loop Returns connector and tell me if anything needs attention before I build this report."
loop_returns"List every return still sitting in 'review' from the last two weeks β customer, dollar total, and how long it's been waiting, oldest first."
loop_return_get"Pull the full picture for order #1244968 β what came back and why, refund vs exchange dollars, and where the package is right now."
loop_return_notes"Read me the internal notes on return #115533968 and summarize the story so far in two sentences."
loop_shipments"Of everything mailed back in the last 10 days, how much is in transit vs delivered β and which delivered packages carry the highest value, so receiving checks them first?"
Live Jul 23: 579 packages in the window β 53 in transit, 16 delivered.
loop_returns_summary"Break June and July returns down by outcome, month by month β is our exchange share improving or are we still refunding most of it?"
Live Jul 1β23: $184,283 return value, $97,498 refunded.
get-shop-info"Before we schedule the weekly report β confirm which Loop shop we're connected to and what timezone it runs on."
read-return-policies"Summarize our current return policy β windows, allowed outcomes, and anything unusual β as a short brief I can share with the new CX hire."
list-returns"Search everything tied to stacey.cohen@ β how many returns has she made all-time, what states are they in, and do any carry a fraud-risk marker?"
get-return"Open the full interactive card for return #115525964 β I want to review it and maybe act from there."
get-return-timeline"Walk me through the timeline of return #115541367 β where exactly did it stall, and did we or the customer cause the delay?"
list-fraud-categories"What fraud categories does Loop define for us, and give me a one-liner on when each applies."
list-allowlist"Who's on our returns allowlist right now? Group them by how they got there if the entries say."
list-blocklist"Show the blocklist and highlight anyone added in the last 90 days so we can sanity-check recent calls."
list-workflows"List our Loop workflow rules by priority and mark which are active β I want to see the automation we're actually running."
get-workflow"Take our highest-priority active workflow and explain in plain English what it does, what triggers it, and who it affects."
add-return-note"Add a note to return #115533968: 'Customer emailed photos Jul 23 β clasp damage confirmed, approved for refund per Melinda's Lifetime Guarantee.'"
flag-return"Flag return #115519374 for review and note why: third high-value refund request from this address this month."
Reversible β flagging asserts nothing, it just calls for human eyes.
unflag-return"That one checked out β unflag return #115519374 and add a note that the address matched a verified gift order."
process-return"Process return #115533968 β but first show me the refund amount and items so I can confirm."
Irreversible; money moves.
reject-return"Reject return #115460305 β the piece came back visibly worn beyond policy. Draft the note documenting why before you do it."
Irreversible.
cancel-return"The customer replied she's keeping the necklace β cancel return #115507404 and note the conversation."
Irreversible.
close-return"Close return #115498193 without warehouse action β it was approved as keep-item, nothing is coming back."
Irreversible.
report-fraud"This one's confirmed β report fraud on return #115541367 with the right Loop category, and put the evidence summary in the comment."
Permanent, once per return, confirmed abuse only. Unsure? Flag instead.
add-to-allowlist"Add our stylist partner sara@agency.com to the allowlist with a note β her sample pulls keep tripping the automation."
remove-from-allowlist"That agency partnership ended in June β remove pulls@old-agency.com from the allowlist."
add-to-blocklist"Blocklist buyer2291@β¦ β four confirmed fraudulent claims this quarter; reference the fraud reports in the note."
remove-from-blocklist"Support resolved the misunderstanding with dana@β¦ β take her off the blocklist and note the ticket number."
Where it gets fun: Loop + Shopify + shiphero26 + Gorgias + Klaviyo + ShopMy + QuickBooks + Asana in single asks. POWER deep analysis Β· AUTOMATION make it a scheduled task. Anything that sends or changes is drafted for your approval first.
"Find our sizing leak: pull this quarter's returns for every size-related reason, rank the top 10 styles by dollars lost, pull each one's Shopify sales so we see return rate not just volume, check whether its exchanges actually shipped in shiphero26 β then draft one Asana task per style for the PDP team with the specific size guidance to fix."
Loop + Shopify + shiphero26 + Asana Β· the $79.9K/July problem, turned into tickets.
"CX morning triage: for every open Gorgias ticket mentioning a return or refund, look up the live return status and tracking, note customers whose packages already delivered but aren't processed yet, and draft a personalized reply per ticket β flag any customer with 3+ lifetime returns for a closer look."
Gorgias + Loop Β· drafts only, you send.
"Exchange integrity audit: take July's Loop exchanges, match each to its EXC- order in shiphero26, and give me three lists β exchanged and shipped, exchanged but stuck in the warehouse (with the blocker), and exchanges Loop shows that never reached ShipHero at all."
Loop + shiphero26 Β· catches revenue-saving exchanges quietly failing.
"Creator risk report: for the styles our top 10 ShopMy creators drove this month, compute each style's Loop return rate against its Shopify sales; flag anything over 15%, break those down by reason β and if a creator favorite is both selling fast AND coming back for sizing, draft the Klaviyo size-guide email for that product's buyers."
ShopMy + Shopify + Loop + Klaviyo.
"Damage forensics: every return citing 'damaged when it arrived' in the last 60 days β pull the original shipment from shiphero26 to see the carrier, service level, and ship date; tell me whether damage clusters by carrier, by style, or by week; and total what damage is costing us in refunds."
Loop + shiphero26 Β· is it packaging, product, or the carrier?
"Finance tie-out: reconcile July's Loop refund totals week by week against QuickBooks β list any week where they drift by more than $500, and for the worst week, list the individual returns so accounting can trace it."
Loop + QuickBooks.
"Fraud-ring sweep: list returns marked high fraud risk this month, group by customer and shipping address to find repeats, pull each repeat customer's lifetime Shopify spend so we don't torch a VIP by mistake β then walk me through them one by one and I'll say flag, blocklist, or clear."
Loop + Shopify Β· analysis is automatic, every action waits for your call.
"Every Monday 8am: build the returns pulse β last week's volume vs the week before, top 3 reasons with $, exchange-vs-refund share, flagged count, and the biggest single return β post the draft to #cx and hold the exec email until I approve."
Scheduled task Β· Loop + Slack + Gmail.
"On the 1st of each month: assemble the full returns report β reasons, outcomes, month-over-month trend, worst styles, fraud summary β as a branded one-pager, save it to the leadership Drive folder, and draft the email to leadership."
Loop + Canva + Google Drive + Gmail Β· MM lavender and all.
"Daily at 9am: any style that crosses 10 sizing-related returns this month gets an Asana ticket for the PDP team with its numbers attached β never duplicate a ticket for the same style twice in one month."
Loop + Asana Β· the sizing leak, watched while you sleep.
In progress: open (awaiting items/processing) Β· processing Β·
review (needs a human before it proceeds).
Finished: processed Β· closed Β· cancelled Β· expired (never completed in time).
Fun fact: inside Loop, a return is literally called a "Loop."
refund money back Β· credit store credit / gift card Β·
exchange replacement piece Β· upsell exchanged up to something pricier β plus combos like
exchange+refund. The refundβexchange mix is the saved-revenue dial.
Methods: mail Β· in_store Β· keep_item (we let them keep it).
Mailed labels move new β pre_transit β in_transit β out_for_delivery β
delivered; you may also see no_label, needs_scheduled, pickup_failed.
Flagging = "a human should look." Reversible, no accusation. Reporting fraud =
a permanent record for confirmed abuse, Loop-defined category, once per return. Unsure? Flag, don't report.
EXC- orders in ShipHero = Loop exchanges.
| Step | Ask Claude |
|---|---|
| 1 Β· Spot it | "Show me returns flagged or marked high fraud risk this week." (Live Jul 23: 83 flagged in 7 days.) |
| 2 Β· Flag early | "Flag return #β¦ and note what looks off." Reversible β it just calls for eyes. |
| 3 Β· Build the 360 | "Show this customer's full returns history, Shopify order history and lifetime spend, and any Gorgias tickets." |
| 4 Β· Decide | Legit β "process it." Not β "reject it." Both irreversible β Claude previews and waits for your yes, every time. |
| 5 Β· Confirmed only | "Report fraud with the right category" (Claude pulls Loop's live category list β never guesses). Repeat offender β "add them to the blocklist." |
Fraud rarely announces itself β it shows up as patterns: the same address, the same claim, the same timing. These prompts are the detective kit. They're all reads, so anyone can run them; acting on what you find (flag β investigate β reject / blocklist / report) follows the playbook in section 7. Golden rule: a pattern justifies a flag, never a fraud report. Reports are for confirmed abuse only.
"Across the last 90 days, group returns by customer email β who has 3 or more? For each, show total refund dollars, their reasons, and whether any return carries a fraud-risk marker."
"Group this quarter's returns by shipping address and flag any address used by more than one customer name β classic sign of one person running multiple accounts."
"Find every return or refund request claiming the item never arrived in the last 60 days. Which customers, which carriers, and does any customer appear twice?"
"Show customers whose refund total this month is over $500 or triple their prior monthly average β sorted by dollars, with each return listed."
"List high-value returns (over $200) opened within 3 days of the return-window deadline, with reasons β the classic wear-it-once-and-return timing."
"Who has received more than one keep-item approval in the last 6 months? Total the dollars we've written off per customer."
"Cross-check this month's returns against Shopify: flag customers whose first-ever order was returned within a week of delivery, and show each one's order value β a common stolen-card pattern."
Loop + Shopify.
"For returns marked received but disputed by the warehouse this month, pull the shiphero26 receiving record and compare what we expected back vs what was actually checked in."
Loop + shiphero26.
"Of returns marked high fraud risk in the last 60 days, how many did we end up processing anyway, and what did those refunds total? List them so we can review the calls."
Claude can bring its other powers β web search, cross-connector joins, pattern research β to bear on a case. Point them at addresses, email domains, tracking numbers, and known scheme patterns: facts about the transaction, not dossiers about the person.
"For the shipping address on return #115519374, search the web: is it a freight forwarder, a commercial mail drop, a reshipping warehouse, or a vacant/for-sale property? Then check how many other customers in Loop and Shopify have shipped to that same address."
Web search + Loop + Shopify Β· reshipper addresses are the #1 stolen-card tell.
"Take the email domains behind this month's high-fraud-risk returns and check the web for which are disposable/temporary email providers β then list which risky returns used a burner domain."
Web search + Loop Β· burner emails cluster hard around fraud.
"The customer says the return was mailed weeks ago. Pull the label's tracking from Loop, verify its real scan history, and compare against what shiphero26 receiving shows β did a package actually move, and did it weigh in?"
Loop + web + shiphero26 Β· fake-tracking and empty-box claims fail this test.
"This case looks like organized refund abuse: search the web for current refund-fraud schemes hitting jewelry and fashion e-commerce (fake-tracking refunds, professional 'refunders,' wardrobing rings), and tell me which known scheme our pattern most resembles and what evidence usually confirms it."
Web search Β· learn the playbook the fraudsters are using β no personal data involved.
"Build the case file for return #115541367: complete Loop history and timeline, every past return by this account, Shopify order history and lifetime spend, Gorgias conversations, shiphero26 shipment + receiving records, address and email-domain checks from the web β assembled as one evidence summary with your read on it, for my decision."
Six systems, one document Β· you decide, Claude documents.
"Run my fraud sweep: repeat returners, shared addresses, 'never arrived' claims, and risky-but-processed returns for the last 30 days β cross-check the worst offenders' lifetime Shopify spend so we don't torch a VIP, then walk me through them one by one and I'll say flag, blocklist, or clear."
Loop + Shopify Β· analysis automatic; every action previewed and waits for your yes.
| Topic | What to know | Status |
|---|---|---|
| Reads for everyone | The analytics pipe is read-only by construction β zero write code, verified by a test that fails the build if a write verb ever appears. Loop's API has 90 write operations; this pipe contains none. | Safe |
| Writes, gated 3Γ | Actions require your personal Loop login and the right Loop role and a yes to Claude's preview. Reject / cancel / close / process are irreversible; everything is in Loop's audit trail under your name. | Role-gated |
| Who signs in | Analytics pipe: MM Google accounts only, enforced server-side. Action pipe: your Loop account only. | Locked down |
| Secrets | The shared read key lives only as an encrypted Cloudflare secret (2-minute rotation). Your Loop session is yours alone β no shared write credentials exist anywhere. | Server-side |
| Date windows | Analytics queries: β€120 days each, no dates = last 24h; 2β4 week windows are the sweet spot at our volume (~60 returns/day) β Claude splits longer asks automatically. Search "last N days" filters count from a return's last update, not creation. | By design |
| Who sees what | Any signed-in MM user can read customer names/emails and refund amounts. Treat exports as confidential. | Handle with care |
| Counting quirk | Reason-level numbers count returned line items; state/month numbers count returns. Claude labels which is which β ask if unsure. | Know it |
| Data quality | Answers mirror Loop exactly β a wrong number in Loop is faithfully reported. High-stakes numbers: ask Claude to show its work. | Verify |
ok: true = healthy. Browser check: loop-mcp.melindamaria.ai/health.| Date Β· version | What changed |
|---|---|
| Jul 23 Β· guide v3.0 | Full redesign: one-connector framing, capabilities by user type, one prompt per tool (all 28), power-prompt showcase, security infographic. Corrected the stock toolkit count to 10 read Β· 12 write (previously misstated as 11/11). |
| Jul 23 Β· loop-mcp v0.1.2 | Hardening from live use: auto-retry on gateway timeouts, smaller analytics pages, compact + filterable shipments. Acceptance: 1,320 returns / $184K summarized clean. |
| Jul 23 Β· v0.1.1 | Deployed to loop-mcp.melindamaria.ai; fixed a Workers-only fetch quirk found in live verification. Connected + verified with real data. |
| Jul 22 Β· v0.1.0 | Built & tested from scratch (research β 6 tools β tests) after a full study of Loop's API (134 operations) and MCP toolkit. Dual-pipe architecture decided. |
A live artifact is a small interactive page Claude builds and saves for you — a dashboard, tracker, or report that pulls fresh data from this connector every time you open it, instead of being a frozen snapshot. In a Cowork session, just ask: “create a live artifact that…” — it lands in your Artifacts list, reusable forever, and its charts and tables refresh themselves on open.
Sharing with the team: artifacts are saved to your Claude account. Three easy ways to share: (1) ask Claude to export a snapshot (PDF, image, or .xlsx) and drop it in Slack; (2) send a teammate the prompt that built it — their copy fills with live data through their own connector sign-in; (3) for a permanent all-team dashboard, that’s the MM Intelligence project on the AI roadmap.