Guide · 3PLs and fulfillment providers
The calls a 3PL fields all day are not hard questions. Where is my order. Did it ship. What do you have of my SKU. Why did it go out of Reno. They are hard only because someone on your side has to stop, log in, find the answer, and read it back. With the InfiPlex MCP server, you issue each client a read-only API key scoped to their instance, they connect the assistant they already use, and those questions get answered without you. Onboarding their marketplace accounts becomes a link you send. And because every key is yours to shape and revoke, the client sees exactly what you allow and nothing else.
The problem
A 3PL with thirty clients answers the same six questions hundreds of times a week. Each one costs a few minutes of someone who should be receiving or picking. Client portals were supposed to fix this and mostly moved the problem: the client has to learn your portal, remember another login, and find the screen. An agent inverts it. The client asks in the tool they already have open, in their own words, and the answer comes from your instance through a key you control.
| What the client asks their agent | Tool it calls on their key | What you no longer do |
|---|---|---|
| "Where is order 7719?" | order_search then order_get |
Look it up and read back the tracking |
| "Did anything of mine ship today?" | order_search with today's window and fully_shipped 1 |
Run a report and email it |
| "How much of SKU 4400 do you have, and where?" | inventory_get per warehouse |
Open the inventory grid |
| "Why did my order go out of Reno instead of Columbus?" | order_warehouse_routing_explanation_get |
Reconstruct the routing decision by hand |
| "What of mine is running low?" | inventory_restock_list |
Forward the restock email |
| "Did my Walmart orders come in overnight?" | log_types_get, log_search |
Check the log and reply |
More client-side prompts are on the prompts page; every read prompt there works on a client key.
The client key
The MCP server publishes a tool to an agent only when the key carries that tool's scope. A client key with read scopes only exposes read tools; the write tools do not exist for that agent, so the client's assistant cannot receive, ship, reprice, or change a setting no matter what it is asked. Credential settings return only whether a value is set, never the value.
Tools > Admin API > Create New API Key on the client's instance. Name it for the client so the API log attributes activity. Check Read Orders, Read Inventory, Read Warehouse, Read Log. Leave every Write scope unchecked.
MCP Enabled: Yes. MCP Allow Setting Writes: No. MCP Allow Arming: No. Save. The key page now shows the Server URL and header the client needs.
The client's instance URL with /mcp and the key as an api-token header. The set-up guide has per-client steps for Claude, Claude Code, Codex, and generic clients; send them that link.
Client leaves, key compromised, staff change on their side: disable the key. Their agent disconnects on the next call. Nothing else on your instance is affected.
Clients on their own InfiPlex instance get a key on that instance and see only that instance. That isolation is the same one the 3PL Connector already gives you; the agent simply inherits it. Full detail on what each key setting unlocks is in how AI agent writes are protected.
Onboarding
The slowest part of onboarding a new client has always been getting their Amazon, Walmart, Etsy, TikTok Shop, or Shopify account authorized without asking for their credentials or walking them through a screen share. Your agent handles it in one call.
The link authorizes; it never exposes a token to you or the agent. Repeat per channel. Then create their warehouse record, set their bin locations, and issue their client key, all from the same conversation on your own key.
Your side
On a warehouse key (read scopes plus Write Orders and Write Inventory), your agent receives, records tracking, sets bin locations, and answers for every client instance you connect it to. Each channel-visible write stops at a preview you approve in chat.
Same key, one paragraph of standing instructions, and a cron line: the agent applies the night's receiving files and records the carrier export before your team arrives, then writes a per-client report. The autonomous fulfillment agent guide has the complete recipe, including a read-only dry run.
Connect one assistant to several client instances, each on its own key. Ask "what is unshipped across all clients" and it asks each instance. A mistake in one client's file cannot touch another's data because the keys do not cross.
Every call runs under a named key, so the API log shows per-client agent activity. If you bill for service levels, you have the record of what was asked and what was done.
The math
Count the client questions your team answers in a week and multiply by the minutes each one takes. For most 3PLs that number is a part-time position. A client key costs nothing, takes two minutes to issue, and the client connects it in five. Onboarding a new client's channels drops from a scheduled call to a link. And the client experience improves: they get answers at 9 PM on a Sunday from the assistant on their phone, not a ticket answered Monday. The cost of one connection page covers the platform economics; this is the service-desk side of the same argument.
Frequently asked
No. A key is issued on one instance and the MCP endpoint is on that instance's subdomain. A client's agent connects to their instance only. Clients sharing a single instance would need your judgment on scopes; the standard model is one instance per client, as with the 3PL Connector.
Not on a read-only key. Write tools are not published to a key without write scopes, so there is nothing for the agent to call. If you want a client to be able to add order notes or set additional info fields, add Write Orders and they gain those; channel-visible writes such as tracking would still stop at a preview the client confirms.
Any MCP-capable assistant: Claude on claude.ai or Claude Desktop, Claude Code, ChatGPT through Codex, Cursor, Windsurf, or a custom agent. No InfiPlex login, no software from you. They paste the URL and header you send and ask their first question.
No. The MCP server is included with API access on every InfiPlex plan. Keys are unlimited. The client's AI vendor bills their model usage.
Every MCP call becomes an API request under the client's key, so your API activity shows it by key exactly like any integration. Name keys for clients and the log reads as a per-client history.
Yes, if you give their key Read Settings; marketplace_connect_link_get sits under the settings read scope and only issues a link. Most 3PLs keep that on their own key and send the link, but either works.
Resources
Bring the client who calls you most. We will connect an agent to their instance on a read-only key and answer their last three questions, tool by tool.
Open any client instance, Tools > Admin API, create a read-only key with MCP Enabled, and send the client the set-up guide. Questions: Contact us, email info@infiplex.com, or call 888‑770‑0857.
Order ids, SKUs, client names, and warehouse names are examples. Claude is a trademark of Anthropic, PBC. ChatGPT and Codex are trademarks of OpenAI. Model Context Protocol is an open standard.
InfiPlex provides eCommerce companies, warehouses, and 3PLs with a robust platform to manage their sales channels.
©2026. All Rights Reserved. Privacy Policy | Terms & License