It's 7 a.m. on Monday, and your analyst is building the same list she builds every week: which dispensaries haven't reordered in 30 days. Export the sales report, paste it into a spreadsheet, filter by last order date, match it against the account list. Forty-five minutes, every Monday. That's 39 hours a year spent on one list.
Your tech lead says there are three ways to stop doing it by hand: use the API, connect an MCP server, or have a coding agent run it from the command line. All three are real options. They're built for different jobs, and picking the wrong one costs you either money or control. Here's how to tell them apart, without writing a line of code.
What's the difference between an API, an MCP server and a CLI?
All three are ways for something outside your ERP to work with the data inside it. The difference is who's on the other end.
- An API (application programming interface) is how one piece of software talks to another. A developer writes code against a documented set of requests, and the code runs the same way every time.
- An MCP server is how an AI assistant talks to your software. MCP, the Model Context Protocol, is an open standard for connecting AI apps like Claude or ChatGPT to outside systems. Instead of code, you ask in plain English, and the assistant picks the right tool.
- A CLI (command-line interface) is a text window where someone types commands. Developers use it, and so do coding agents that run commands on a developer's machine.
Think of your warehouse. The API is the loading dock: trucks arrive on a schedule with paperwork in a fixed format. The MCP server is the front desk: someone walks up, asks a question in their own words, and gets an answer from staff who know where everything is. The CLI is the back office terminal, where one person with the right login types exact instructions.
MCP isn't new anymore. Anthropic open-sourced it in November 2024, and a growing list of cannabis software vendors now offer an MCP server. So the useful question isn't what MCP is. It's which door each job should go through.
When is an API the right tool?
Use the API when software talks to software, or when a job runs the same way every time.
- Systems that need to stay connected: your CRM, a data warehouse, a BI dashboard, middleware like Zapier or n8n, or a partner's app. Those systems speak code, not English.
- Scheduled and high-volume jobs: a nightly inventory export, a sync that runs every 15 minutes, thousands of records at a time.
- Anything that must give the same answer every run: if the Monday reorder list should follow one exact rule, code it once and it never drifts.
Here's how it works in Distru, a cannabis ERP with a built-in B2B wholesale marketplace. The Distru API is a REST API that returns JSON. An admin issues API keys, and each key expires after a year, so nothing stays open forever. Webhooks can tell your system when something changes instead of making it check over and over. Your account team turns API access on for your account first, and the developer hub has the reference, authentication guide and OpenAPI spec.
The trade-off: someone has to build it and keep it running. That's fine for a job you'll run thousands of times. It's overkill for a question you'll ask once.
When is MCP the right tool?
Use MCP when a person has a question nobody built a report for, or wants to work across several tools at once.
"Which dispensaries haven't reordered in 30 days, and what did they buy last time?" With MCP, you type that into Claude or ChatGPT and get the list back without an export or a spreadsheet. Next week you might ask something different, and you don't need a developer to change anything. You can also connect ChatGPT to your cannabis ERP and, if it's connected too, your email, so the follow-up drafts sit right next to the list.
In Distru MCP, you sign in with your own Distru login. Claude or ChatGPT sees exactly what your role in Distru allows, and nothing more. If your role can't view costs, the assistant can't either. You set it up in Claude's or ChatGPT's connector settings; our post on connecting Claude and ChatGPT to Distru walks through it.
The trade-off: an AI model reads your question and decides which tools to call, so two people asking slightly different questions can get slightly different answers. For exploration, that flexibility is the point. For a number that goes into a board deck, check it against the report.

Where does a CLI fit for a cannabis business?
For most operators, it doesn't, directly. A CLI is a developer's tool. Where it shows up now is through coding agents, AI tools that run commands on a developer's computer to build and test software.
Research backs up the split. A July 2026 comparison of MCP and CLI for AI agents found the command line more efficient for one person working on their own machine, while MCP covers what a business needs once several people and systems are involved: access control, credentials kept out of the agent's hands, and an audit trail.
If your developer uses a coding agent, it works through the Distru API, and the developer hub publishes plain-text docs written for coding assistants so they can learn the API quickly. In practice, the CLI question for an operator is really an API question: your developer's tools will call the API either way.
What does each one cost you to run?
Every MCP request goes through an AI model, and the model's provider bills by usage. Ask one question, and the assistant may make several calls behind the scenes. That's cheap for a few questions a day and adds up for a job that runs every five minutes.
API calls don't use an AI model at all. Once the code is written, a thousand runs cost little more than one. So the rule of thumb is simple: questions go to MCP, repetition goes to the API. If your team asks the same question every Monday, that's the signal to have someone turn it into an API job.
Should you build your own integration instead?
Your tech lead may say they can build all of this themselves, and they're probably right. Calling an API isn't hard. Keeping it working is the hard part.
Metrc rules vary by state and change often. Accounting systems change their integrations. Retailers change what they need on an invoice. Every change means someone updates your code, tests it and fixes what broke. Before you build, ask who maintains it in a year, and what happens to your Monday list when they're on vacation.
A reasonable split: let your ERP's maintained API and MCP server carry the Metrc and accounting plumbing, and point your team's time at the analysis only your business can do.
Who approves what the AI changes?
This is the question operators ask most, and it should be. Reading data through an AI assistant is low risk. Changing data is different.
- Start read-only. Get comfortable with answers before you let the assistant change anything.
- Keep approval on. Whether Claude or ChatGPT asks before it changes anything depends on that app's settings, not on Distru. Set it to ask, and read the preview before you say yes.
- Check results against your own data. When a number matters, confirm it in Distru before you act on it.
Distru adds its own limits. Write tools are rolling out in stages, and no tool creates sales orders yet. Assemblies made through Distru MCP save as pending for your team to review in Distru. Metrc reporting stays in the Distru app, where your team presses sync. For more on setting permissions, see what happens when AI gets access to your ERP.
Which one should your team use?
Run each job through three questions: who's asking, how often, and does it read or write?
- A person, asking something new: MCP. Ad hoc questions, one-off lists, "why did this account drop off?"
- A person, asking the same thing every week: MCP to figure out the right question, then the API to run it on a schedule.
- Software talking to software: the API. CRM syncs, BI dashboards, partner apps, warehouse exports.
- A developer building or testing: their coding agent and CLI, calling the API underneath.
- Anything that changes records: MCP with approval on, or the API with code your team has reviewed. Never an unreviewed loop.
Back to the Monday list. Start by asking for it through MCP for a few weeks. Once the question stops changing, have your developer turn it into an API job that lands in your inbox at 6:55 a.m.
Where this stops working
None of this fixes messy data. If product names are inconsistent or accounts are duplicated, an assistant will find the mess faster but won't clean it up for you. And MCP answers are only as good as the question: vague prompts get vague results, so be specific about dates, accounts and locations.
It also doesn't replace a person who knows the business. The assistant can tell you which accounts stopped ordering. Your sales team still has to know why.
Two doors into the same records
In Distru, the API and Distru MCP open onto the same records 450+ cannabis operators already run on: inventory, orders, invoices and accounts. Your developer uses one door, your ops team uses the other, and both follow the permissions you set. If you're new to MCP, what an MCP server is covers the basics.

If you want to see both on your own data, we'll walk your actual Monday list, not a canned demo. Builders can start on their own at the developer hub.










