The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Packrift MCP listing page.
Production MCP (Model Context Protocol) server for Packrift exact-spec packaging procurement. Hero use case: find the right packaging supply for a given item, SKU, or reorder need, then confirm live price, inventory, shipping, and cart handoff.
2025-04), store packrift.myshopify.comPOST /mcp, GET /mcp (SSE), GET / and GET /start (MCP start page), GET /.well-known/mcp/server-card.jsonllms-install.md for remote MCP client configuration.The public Packrift endpoint is hosted at:
Fast start page for developers, agents, and directory reviewers:
Directory and partner handoffs can use tracked start links without changing the MCP endpoint. Use a lowercase source slug with letters, numbers, and underscores:
The start page also renders source-specific copy controls when it receives a source:
Those handoffs record aggregate install-intent telemetry: /r/config/{source} fetches show up as mcp_tracked_config_fetches, /r/install/{source}/{target} opens show up as mcp_install_intent, and copy controls record mcp_install_copy by source and target so partner and directory handoffs can be evaluated before downstream cart events appear.
Custom partner, campaign, directory, and agent-workflow slugs are allowed without code changes as long as they match ^[a-z0-9_]{2,64}$. Examples: mcpservers_org, agency_partner, browser_agent_demo, newsletter_mcp.
When a directory or agent host wants a source-attributed config file instead of an HTML start page, use:
Tracked config fetches are exposed by source in https://mcp.packrift.com/ai/mcp-usage-snapshot.json.
For a target-specific tracked install action, use:
Common targets include generic_streamable_http, stdio_mcp_remote, claude_code, codex, claude_desktop, cursor_windsurf_vscode, and cline.
MCP clients that support remote HTTP or Streamable HTTP servers can add Packrift with:
Cline users should use the tracked Cline target, which returns Cline-native streamableHttp config:
Hosts that only accept local stdio MCP commands can use the tracked stdio_mcp_remote target. It runs npx mcp-remote as a thin bridge and still forwards every call to the hosted Packrift endpoint:
Use the hosted endpoint above when possible. It requires no buyer-side API key and exposes the current 15-tool Packrift exact-spec commerce surface.
The hosted Glama connector should remain the primary Glama traffic target. The separate Glama source server listing can use this repository for release and quality checks, but it should still point users at the hosted no-auth MCP endpoint above.
Dockerfile starts the Packrift MCP server on PORT=8787.glama.json is the source-listing maintainer claim file.smithery.yaml gives Glama/Smithery-compatible container metadata with an
empty config schema, so source-listing release checks do not infer a required
buyer token.SHOPIFY_PACKRIFT_TOKEN still exposes MCP
discovery: tools/list returns the current 15-tool surface and
resources/list returns the public AI/MCP resources.SHOPIFY_PACKRIFT_TOKEN is not required for MCP discovery, directory
scanning, or hosted connector use. It is required only for self-hosted live
Shopify-backed catalog, pricing, inventory, shipping, and cart tool calls.https://mcp.packrift.com/mcp.The public hosted endpoint above is the primary integration path and the
canonical server.json manifest is remote-only. A container image is still
published for local development and environments that explicitly require a
self-hosted package-style install surface:
You can also build the image from this repository:
Optional environment variables:
PORT defaults to 8787SHOPIFY_STORE_DOMAIN defaults to packrift.myshopify.comSHOPIFY_API_VERSION defaults to 2025-04STOREFRONT_DOMAIN defaults to packrift.comSHOPIFY_PACKRIFT_TOKEN is optional for MCP introspection and directory scanning, but required when self-hosting live Shopify-backed catalog, pricing, inventory, shipping, and cart toolsThe container uses an in-memory cache instead of Cloudflare KV. It is intended
for discovery and client testing; the production server remains the Cloudflare
Workers endpoint at https://mcp.packrift.com/mcp.
llms.txt: https://mcp.packrift.com/llms.txtllms-full.txt: https://mcp.packrift.com/llms-full.txt^[a-z0-9_]{2,64}$; partner-specific slugs are allowed without pre-registrationPACKRIFT_EXTERNAL_ACTIVATION=1 curl -sS 'https://mcp.packrift.com/ai/mcp-activation-wave-runner.sh' | bashPACKRIFT_EXTERNAL_ACTIVATION=1 PACKRIFT_ACTIVATION_WAVE_SCOPE=full curl -sS 'https://mcp.packrift.com/ai/mcp-activation-wave-runner.sh' | bashcreate_cart_url, replaces the cart button with the fresh measured /r/cart/1066 URL, and still does not place an order.Notify IndexNow about the existing source-specific MCP eval-pack, start, install, first-run, and activation URLs:
The script preflights https://mcp.packrift.com/ai/mcp-source-activation-sitemap.xml, writes artifacts under outputs/source-activation-indexnow/, and submits only existing hosted MCP URLs. It does not create a new CLI, buyer surface, checkout path, or directory submission.
Tools are framed around exact-spec procurement, not generic browsing. Use find_packaging_for_item when the buyer has item dimensions or a fit question; use SKU and exact-spec tools when the buyer is replenishing a known product.
| Tool | Purpose |
|---|---|
find_packaging_for_item(item_length_in, item_width_in, item_depth_in, item_weight_lb, use_case) | Hero. Item L/W/D + weight + use case -> ranked packaging SKUs that fit. Use for smallest-fit, box-vs-mailer, and Uline-by-size style questions. |
search_products(query, limit?) | Keyword fallback when dimensions are unknown, such as kraft tape, bubble mailer, starter kit, or weather-resistant labels. Cached 5 min in KV. |
get_product(handle) | Full product detail including variants, dimensions/metafields, weight, stock, and product URL. |
get_pricing(variant_ids[], quantity?) | Live unit price and line total before purchase handoff. variant_ids must be numeric Shopify variant IDs encoded as strings. Never cached. |
check_inventory(variant_ids[]) | Live inventory check before recommending or building a cart. variant_ids must be numeric Shopify variant IDs encoded as strings. Never cached. |
get_shipping_estimate(zip, country, items[]) | Carrier rates for a chosen cart via Shopify draftOrderCalculate. |
get_cart_handoff_candidates(limit?, family?, sku?) | Priority curated SKUs with ready create_cart_url arguments, SKU/product/reorder/quote/cart links, and required live-confirmation sequence. |
create_cart_url(items[], discount_code?, ref?) | Final cart handoff. Builds a GA4-visible Packrift cart landing URL plus the final packrift.com/cart/... permalink with ref=mcp and AI-commerce attribution fields. |
prepare_purchase_handoff(sku, quantity?, buyer_confirmed?) | Fast exact-SKU path. Confirms the product, live price, and inventory, then returns a tracked cart URL only when buyer_confirmed=true. |
compare_alternatives(requested_spec, family?, competitor_reference?, limit?) | Ranks curated Packrift alternatives for open-ended buyer requests, including competitor-style packaging specs. |
pack_calculator(item, padding?, use_case?, limit?) | Calculates protected inside dimensions and returns fitted box or mailer candidates with void-fill guidance. |
inventory_status(variant_ids?, sku?, handle?, quantity?) | Live total and location-level Shopify inventory status for exact SKUs, handles, or variant IDs. |
get_reorder_link(sku, handle?) | Reorder URL, product URL, and copy-procurement-spec text for a catalog SKU or handle. |
get_bulk_quote_link(requested_spec, family?, sku?, quantity?) | Tracked bulk quote URL for no-exact-match, large-volume, or procurement-review workflows. |
explain_no_exact_match(requested_spec, missing_or_mismatched_fields?) | Explains why a nearby product should not be presented as exact, then routes to safe next actions and quote recovery. |
The server advertises prompts and resources in addition to tools.
Prompts:
find_exact_packaging_specfit_item_then_prepare_cartreview_cart_handoff_candidatesreorder_packrift_skuprepare_cart_handoffrequest_bulk_quote_for_no_matchcopy_procurement_specfind_box_by_lwhfind_label_by_size_material_printerResources include the AI-approved product corpus, per-SKU markdown/JSON records, purchase-path JSONL, MCP cart handoff candidates, MCP cart activation playbook, measured handoff directories with cart routes, exact-spec family files, no-exact-match policy, crawler-safe purchase paths, Shopify-native UCP commerce surface, UCP starter catalog, storefront builder kit, and flat storefront import feed for curated packaging/shipping-supplies shelves.
Packrift is optimized for exact-spec packaging procurement: agents should not treat nearby sizes, materials, colors, adhesives, strength ratings, or case counts as exact substitutes. If there is no exact AI-approved match, route the buyer to reorder, no-match explanation, or bulk quote recovery instead of forcing a substitute.
Current paid ChatGPT / product-card order families that are represented in the public corpus:
1066: 10 x 6 x 6 ECT-32 kraft corrugated boxes, 25/bundleMFL1295: 12 1/8 x 9 1/4 x 5 white self-seal literature mailers, 50/packLL251WR: 2 5/8 x 1 weather-resistant polyester laser labels, 3000/caseInstall Node 24+ and the deps:
Use /Users/farhan/Downloads/packrift-mcp as the canonical production worktree. The older
/Users/farhan/Downloads/packrift-mcp-server checkout is a legacy duplicate and must not be
used for Worker deploys because it can overwrite mcp.packrift.com with stale routes/tools.
Local secrets belong in .dev.vars (gitignored). Do not paste real tokens into
this README, issues, directory submissions, screenshots, or third-party support
threads.
Run the server:
Smoke-test the MCP endpoint with curl:
Type-check:
Smoke-test the hosted cart handoff path without creating an order or collecting payment:
This synthetic check calls the current hosted MCP endpoint, verifies
get_cart_handoff_candidates -> get_product -> get_pricing -> check_inventory -> create_cart_url, and checks the MCP cart landing route. It writes JSON and
Markdown evidence under outputs/mcp-cart-handoff-smoke/. Add
-- --sku LL251WR --qty 1 to test another AI-approved SKU. Use
-- --verify-final-cart only when intentionally checking Shopify's final cart
redirect.
GitHub Actions also runs the hosted smoke checks every six hours and on manual
dispatch through .github/workflows/mcp-live-smoke.yml. The scheduled job
checks both 1066 and LL251WR, then uploads smoke and capture evidence
artifacts. Run npm run check:distribution for the broader directory and
registry monitor; it exits nonzero only when a tracked surface has a true
fail status.
Refresh the llms-full.txt priority SKU block from recent GA4 item activity joined to the AI-approved catalog and live Shopify inventory:
The generated public block must not include local paths, internal notes, or non-public campaign details. A private JSON report is written under outputs/llms-full-priority-skus/.
To run the whole refresh path, pull fresh GA4 reports, publish the public llms-full.txt override into Workers KV, purge the public agent URLs, and verify the live surfaces:
This updates the static-override:llms-full.txt KV key. The Worker serves that override after the current deploy, so future priority-SKU refreshes do not require deploying unrelated Worker source changes.
Check crawler-facing llms-full.txt availability across six AI user agents:
This writes JSON and Markdown evidence under outputs/static-availability/. The funnel snapshot also runs this check and reports failure rate, 5xx rate, content validation, and p95 latency.
Refresh the full MCP funnel proof, including first-party Shopify order stitching for MCP cart attributes:
The snapshot reads the token-protected https://mcp.packrift.com/admin/mcp-orders
endpoint, scans recent Shopify orders for MCP cart attributes such as
packrift_mcp_key, packrift_ai_id, and utm_source=chatgpt-mcp, and reports
attributed order count and revenue beside GA4 cart-landing evidence. The admin
endpoint requires MCP_STATS_TOKEN and does not expose customer PII.
Publish the sanitized GA4 visitor proof from that local snapshot to the public Worker surface:
This updates https://mcp.packrift.com/ai/mcp-ga4-funnel-proof.json and .md
without exposing local paths, raw CSV rows, order rows, or credentials. The
public funnel snapshot also reads this proof for the thousands-of-qualified-
visitors gate when it is available.
Refresh MCP directory evidence:
The distribution check reports which public MCP directories are current, stale, blocked, or failing. The submission pack turns stale rows into copy-ready listing fields, live proof URLs, and the next refresh action for each directory.
Check the all-agent capture hub:
This verifies that https://mcp.packrift.com/ai/all-agent-capture.json and
.md, https://mcp.packrift.com/ai/mcp-adoption-kit.json and .md,
https://mcp.packrift.com/ai/mcp-install-matrix.json and .md,
https://mcp.packrift.com/ai/mcp-usage-snapshot.json and .md, and
https://mcp.packrift.com/ai/mcp-agent-adoption-progress.json, .md, and .html, and
https://mcp.packrift.com/ai/mcp-buyer-use-cases.json, .md, and .html, and
https://mcp.packrift.com/ai/mcp-cart-activation.json, .md, and .html, and
https://mcp.packrift.com/ai/mcp-workflow-gallery.json, .md, and .html, and
https://mcp.packrift.com/ai/browser-agent-bridge.json and .md, and
https://mcp.packrift.com/SKILL.md, and
https://mcp.packrift.com/ai/mcp-directory-refresh.json and .md, and
https://mcp.packrift.com/ai/mcp-directory-submit-actions.json and .md, and
https://mcp.packrift.com/ai/mcp-reviewer-activation.json and .md, and
the activation command center at https://mcp.packrift.com/r/activate,
the source activation queue HTML at https://mcp.packrift.com/ai/mcp-source-activation-queue.html,
the activation wave and guarded full-source capture runner at https://mcp.packrift.com/ai/mcp-activation-wave.json,
the selected external activation task queue at https://mcp.packrift.com/ai/mcp-external-activation-brief-tasks.jsonl,
the compact selected-task exports at https://mcp.packrift.com/ai/mcp-external-activation-brief-tasks.jsonl?compact=1
and https://mcp.packrift.com/ai/mcp-external-activation-brief-tasks.csv?compact=1,
the browser runner at https://mcp.packrift.com/r/activate/generic?format=html, and
the shell activation script at https://mcp.packrift.com/r/activate/generic?format=sh, and
https://mcp.packrift.com/ai/mcp-eval-pack.json and .md, and
https://mcp.packrift.com/ai/agent-capture-outreach.json and .md are live,
advertised in resources/list, include the core agent surfaces,
first-five-minute install/test path, copy-paste developer examples, copy-ready install matrix, public usage snapshot,
agent adoption progress board, buyer workflow map, cart activation playbook, workflow gallery, eval pack, browser-agent bridge, directory refresh pack, directory submit-action queue,
source activation command center, reviewer activation handoff, agent capture outreach packet, Browserbase Browse candidate lane, and root SKILL.md, and preserve the
rule that Packrift uses the hosted MCP endpoint instead of a duplicate CLI surface. The outreach build writes
copy-ready stale-directory refresh messages, Browserbase Browse candidate
briefs, proof links, and MCP install snippets under
outputs/agent-capture-outreach/.
The production endpoint is https://mcp.packrift.com/mcp. Use the hosted
endpoint for buyers, directory reviewers, and agent integrations. Only deploy a
self-hosted worker when you are updating the production service or testing a
controlled fork.
When deploying with an authorized Cloudflare session, run:
The server card lives at https://mcp.packrift.com/.well-known/mcp/server-card.json.
cartCreate is a Storefront API mutation, not Admin. The brief asked for cartCreate + cartBuyerIdentityUpdate for shipping rates, but those don't exist on the Admin GraphQL API this server uses. The supported Admin path is draftOrderCalculate, which is what get_shipping_estimate uses. It returns the same carrier-rate data without creating a real order.custom.specN_value metafields where the matching custom.specN_name says "Dimensions" or "Size". Format is human-readable (12 1/8" L x 11 5/8" W x 2 5/8" H). src/dimensions.ts parses mixed fractions and falls back to scanning the title.mailer-boxes collection — that handle does not exist on the live store. We use mailers-envelopes, boxes-mailers, corrugated-boxes, bubble-wrap-foam, cushioning, and ecommerce-fulfillment (verified via collections query 2026-04-29).src/tools/recommend_packaging.ts (COLLECTIONS_BY_USE_CASE).handle in the response is a long opaque JWT-style string — that's how Shopify returns rate handles; pass it through to subsequent calls if needed.{ content: [...], isError: true } per the MCP spec, not as JSON-RPC -3260x errors. Protocol-level errors (unknown tool, bad JSON) do return JSON-RPC errors.Packrift publishes AI-commerce SKU pages for exact-spec procurement agents.
Each SKU page links to the canonical Packrift product page, reorder path, bulk quote path, MCP endpoint, and no-exact-match policy. Use these pages only for AI-approved products and confirm live price and inventory through the MCP tools before checkout handoff.