Building Plain Freight: An AI Agent That Runs A China-To-USA Freight Business

Why I am building a freight company

Most of my work is software: backend systems, cloud infrastructure, and AI integrations for startups and scale-ups. Plain Freight is different. It is a freight forwarder — a business that physically moves boxes from a warehouse in China to a doorstep in the United States — and I am building it because the operation behind it is a near-perfect target for AI agents.

Freight forwarding for small shipments is slow, manual, and full of repetitive coordination: quoting, chasing warehouses, filling customs forms, copying tracking numbers between systems, and answering the same customer questions over and over. That is not work that needs human judgement most of the time. It is work that needs a reliable agent that never forgets a step. So I am building the company and the automation together, in public, at plainfreight.com.

The problem: small importers get ignored

Big freight players are built for big accounts. If you move containers every week, you get a dedicated account manager, good rates, and someone who answers the phone. If you ship anything from a 100 g sample to 2,000 kg at a time, which is the reality for most Amazon FBA sellers, Shopify brands, and independent importers, you are too small to matter. You overpay couriers like DHL and FedEx, or you wrestle with a forwarder portal that assumes you already speak the language of HS codes, incoterms, and customs bonds.

The timing matters too. With the end of the $800 de-minimis exemption, a lot of small importers who used to slip under the customs threshold now face real paperwork and duties for the first time. That is a wall of complexity landing on exactly the people with the least support. Plain Freight exists to take that wall away: one all-in price covering freight, customs, and delivery, quoted in seconds, with the messy parts handled for you.

The promise is deliberately small and clear: ship from China, and don't worry about the rest.

How the agent system works

Under the hood, Plain Freight is not one big AI. It is a team of specialised agents, each owning one part of the operation, coordinated by an orchestrator that knows when to escalate to a human. This mirrors how a real freight desk is organised, which makes each agent easier to reason about, test, and trust.

A customer describes the shipment through the website form or by email. The Customer Agent prices it from the live rate card in seconds; a clean, firm offer can go out on its own, and anything unusual waits for my approval. Once the customer accepts, the Operations Agent lines up pickup with the China warehouse, and the customer pays an ACH or wire invoice only after the cartons are weighed there, never by card. The Customs Agent classifies the goods, assigns the right HS code, and generates the commercial invoice and packing list automatically. Those documents go to whoever files the entry on that lane, and the duties are paid inside the all-in price. A Tracking Agent then watches the shipment and pushes plain-language updates back to the customer, while a Finance Agent invoices and keeps an eye on margins.

An Overseer agent traces the whole system — health checks, every agent, every handoff — scores platform health, and proposes fixes when something drifts. I approve anything touching money, restricted or unusual goods, partner commitments and promises to third parties, usually with one tap on Telegram. The routine work in between, the drafting, chasing, classifying and tracking, runs without me.

The tech stack

The agents are written in Python and run on the Claude API, with cheaper, faster models handling routine classification and stronger models reserved for the genuinely hard calls like ambiguous HS codes. Orchestration runs on AWS Lambda with SQS and SNS for queuing and fan-out, so each agent reacts to events rather than polling.

Data lives in Supabase (Postgres), documents are stored in S3 and generated with WeasyPrint, and the customer-facing site is a Next.js app. Customers reach it by email and through the website, partner coordination with the China warehouse uses WeChat Work, communication with whoever clears the entry goes over email, and my approvals run through the Telegram Bot API. Payment is an ACH or wire invoice, with no cards. None of these are exotic choices — the interesting part is not the individual tools, it is wiring them into agents that can carry a shipment from quote to delivery, with me approving the decisions that matter.

Manual first, automate as I go

I am not trying to launch a fully autonomous platform on day one. That is how logistics startups die — they build elaborate automation for edge cases they have never actually seen. The plan is the opposite: start as a manual freight agent, take real shipments, feel the real pain points, and automate each step only once I understand it.

This is the path Flexport's founder took: Ryan Petersen brokered shipments by hand before writing a line of code. The manual phase is not a detour, it is the research. Every awkward customs question and every warehouse miscommunication becomes a precise specification for the agent that will eventually own that step. Software built on top of real operational knowledge beats software built on assumptions every time.

Where it is now and what is next

Today, the Plain Freight website, instant offers and quote-by-email are live, and real shipments are moving while individual steps get automated. You can get a real quote right now and have a shipment handled end to end.

Next is deepening the agent layer: a sharper HS-code classifier, fully automatic document generation, a customs pre-check that flags common errors before submission, and a landed-cost calculator that includes every fee. After that comes more shipping options — sea LCL and express — and a forwarder rating system built from real delivery data. If you import from China, or you just want to follow a freight company being built in the open with AI at its core, take a look at plainfreight.com.