Fulfil is the harness for your AI agents.
A model can reason. To do work in your business it needs records to read, actions it may take, limits, an approval step and a log. That is the harness. Fulfil has it for your orders, your inventory, your fulfillment and your accounting, so what your team builds is the workflow.
A model cannot work in your business until something gives it records, actions and limits, and building that yourself is the long part.
The model is rented by the month and gets better each quarter. The connectors, the joined data, the permissions and the log are yours to build and to keep working. These are brands on that work.
Months on a build of your own
“We were trying to build our own AI forecasting system. Fulfil launched their Claude integration, and it’s been way better than what we spent four months trying to build ourselves.”
A developer between you and the tool
“The app I built had a working prototype in three weeks, because I was able to do it myself and not have to translate everything I needed to a developer, have that person build it or not build it, and then take another stab at it.”
An agent is as good as the records under it
“The micro apps you use are only going to be as good as the data you have inside your ERP.”
The harness is the part between the model and your business.
You choose the model, and you can change it. Your business is your channels, your warehouses and your bank. Fulfil is the part in the middle that both of them already work through.
Each part is already in Fulfil, because your team works through the same parts.
Orders, inventory, fulfillment, EDI and accounting are one set of records. An agent reads what your team reads.
Before it acts, an agent reads the models, the fields and the methods of your workspace, and searches the help articles.
It creates orders, posts journal entries, applies payments and adjusts inventory through the API that the product uses.
Each person signs in as themselves, and the agent can reach what that person can reach. An administrator chooses who may use MCP.
Write tools are off until you turn them on. In Ask mode the agent shows each change and waits for you.
Each write is recorded with the user, the action and the time, next to the changes that people made.
A micro app runs at your workspace address with the Fulfil login. Each version has a preview, and you can roll back.
A webhook tells your agent that a record changed. An automation rule does the routine part with no model at all.
This is what you build when you build your own, and what is already there on Fulfil.
Both ways use the same models. The difference is how much you build and keep working before an agent does its first useful job.
| Building your own, over separate tools | Building on Fulfil | |
|---|---|---|
| Records | A connector for the store, each marketplace, each 3PL and the accounting system, and a database where you join them. | Orders, inventory, fulfillment, EDI and accounting are already one set of records. |
| One number | Each tool keeps its own count and its own IDs. You write the rules that say which one is right. | One count for each product, and the ledger is posted from the same record. |
| Actions | A tool to write and test for each action in each system. | The actions of the product: create an order, post a journal entry, apply a payment, adjust inventory. |
| Permissions | An API key for each system. What each person’s agent may do is yours to design. | Each person signs in with their Fulfil login, and their agent has their permissions. |
| Approval and log | A review step and a log across the systems, built by you. | Write tools start off. Ask mode shows each change first. Each write is in the audit trail. |
| Tools for the team | Hosting, sign-in and deploys for each dashboard or small app. | Micro apps run inside Fulfil with the Fulfil login, a preview and a rollback. |
| Upkeep | When a channel or a tool changes its API, your connector stops until somebody fixes it. | Fulfil keeps the channel, 3PL and EDI connections working. They are the ones the product runs on. |
| What you build | The harness first, then the workflow. | The workflow: the instructions, the checks and the schedule for the job. |
When a build of your own fits better
Fulfil is the harness because it is the system of record. If your orders and your ledger stay where they are, build over them.
Ad creative, support replies and product content read other systems. An agent for those needs little from an ERP.
A team that already runs its own data platform may want the control of its own harness, and can still read Fulfil through the API.
What stays your work on Fulfil
An agent repeats what the records say. Costs, lead times and counts must be right before you hand a job over.
Say what the job is, how it is done today and when to stop and ask, as you would to a new hire.
You pay for your assistant, and the schedule of a job runs there or in your own runner.
Start with questions, and hand over a whole job last.
Each stage uses the same connection and the same permissions. Stay at one until you trust it.
Connect your assistant and only read.
Paste one address into Claude or ChatGPT and sign in. Write tools stay off. The team asks about orders, inventory and cash, and you see where the records are wrong.
Let it change records, with your approval.
Set the write tools to Ask. The assistant drafts the purchase orders or the bin changes, shows you each one and waits.
Atlas Pet Company moved its parts to a two-bin system this way
Give a repeated job its own screen.
Describe the dashboard or the queue, and Claude builds it as a micro app on live records. Your team opens it with the Fulfil login.
Give an agent a job description and a schedule.
Write the job as you would for a new hire. Give the agent a user of its own with the permissions of that job, and run it from Claude Code on a schedule.
At Atlas Pet Company an agent has done this for about six months
Some jobs need no build at all. Fulfil has agents of its own for defined jobs, such as coding bank and card transactions, first in shadow mode and then up to an amount that you set. See the agents in Fulfil
At Atlas Pet Company, an agent named Kai does the work of a head of operations.
“So I took their job description, dumped it into Claude Code, and said: ‘This person’s gone. We need to replace them with an agent, and we need to build this into Fulfil and Claude Code so it’s all automated and scheduled.’”
Atlas makes dog leashes and harnesses in Colorado with four full-time employees. Three people took the head of operations role and none stayed 90 days. Kai works through Fulfil MCP, and has for about six months. Sam gives his agents full access. That is his choice: on a new connection the write tools are off.
- A customer writes at 1 a.m. to change an address. Kai holds the order, changes it and releases the shipment
- When Kai cannot make a change, it flags the order and leaves it on hold for a person
- An agent keeps the accounting in Fulfil up to date each day
Atlas Pet CompanyDog gear, made in Colorado
Operators build on Fulfil themselves.
Katie PowersSenior Director of Accounting, Caraway

“So I’ve been able to build out a micro app within Fulfil that like refreshes and pulls in the real time balance sheet balances. And I can do what I was doing in Excel in this micro app.”
Will AndrewsVP of Operations, Grunt Style

“I am not a developer. I’m just an operator with a passion for technology because it usually helps my people. In less than 60 days, I built an entire ecosystem my whole company is using.”
ErinRetail Planning & Analysis, Cora
“I think Fulfil has really been good about enabling operators like ourselves to build what we see as useful.”
Results at other brands

4 months spent on an AI forecasting system of its own, before it used Claude on Fulfil
Read the Qure Skincare storyEvery hour the operations team asks Claude about inventory, work orders and orders
Read the Evereden story
4 to 6 hours a week saved across the team, by its own estimate, with a ticket system for invoices built with Claude Code
Read the J. Stockard Fly Fishing storyEverything you can build with, on one set of records.
The ways in for a model.
Each one reaches the same records with the same permissions.
Fulfil MCP
For an assistant in a chat window: Claude, ChatGPT, Grok, OpenClaw.
Fulfil CLI
For an agent in a terminal: Claude Code, Cursor, Codex. It ships with a skill that teaches the agent Fulfil.
API and webhooks
For your own code, and for events that start a job.
Cursor plugin
Rules, skills and commands for a coding agent.
Questions
How building on Fulfil works.
What is an AI harness?
Do I need developers to build AI workflows on Fulfil?
Which AI models and assistants can I use?
Can an agent run on a schedule with nobody at the keyboard?
What can an agent change in Fulfil, and how do I limit it?
When is building our own AI stack the better choice?
Works with Claude and the assistants you already use.
Claude
ChatGPT
Grok
OpenClaw
- Your own agents
An agent reads the orders, the inventory, the purchase orders and the ledger that your team uses, with the permissions you set.
Example
Which orders older than 30 days have not shipped, and what should we do with each?
Nine orders, in three groups. Five wait for TWL-204, which arrives on the purchase order due Thursday. Three are wholesale orders on hold for credit. One has an address that the carrier rejected. I can draft the emails for the five, and I have left the hold on the other four for a person.
Answered from the same records that your team uses