Your agents need write access. Here's how to give it to them safely.

I build AI agents for small businesses. Every client I’ve worked with starts in the same place. We connect an AI to their business data, and it can look things up: find a customer, summarize the week, answer questions about the numbers. That’s useful.

Then, almost right away, they ask the next question: “Can it just do it?” Update the ticket. Log the time. Record the expense. Close out the task.

Of course they ask. Nobody wants a very smart search box. Read-only gets you answers. Write access gets work done.

And that’s where I kept getting stuck.

Why write access is scary

The tools agents write into were built for people editing one record at a time. Most of them keep no history you can actually use. If a person makes a mistake, it’s one record, and someone usually notices.

An agent is different. It can misread an instruction, get stuck in a loop, or be tricked by text hidden in an email it was asked to read. When that happens, it can overwrite or delete hundreds of records before anyone looks. There’s often no way to see what changed, and no way to get it back.

So business owners end up with two bad options. Keep agents read-only and give up most of the value. Or turn on write access and carry a risk nobody can measure. I didn’t want to hand my clients either one.

We already solved this once

Coding is the one place AI agents have really taken off, and I don’t think that’s an accident. When Claude Code or Cursor edits my code, I’m not nervous. Every change goes through version control. I can see exactly what the agent did, and if it’s wrong I undo it in one command.

Version control is what made it safe to let agents write code. Business data has never had anything like it. Your CRM, your ticketing system and your spreadsheets just hold the latest version of each record, and whatever was there before is gone.

I made this argument last week in Your agents need write access. This post is about what I built to fix it.

Meet AvianSuite

AvianSuite is one secure place where the AI you already use, like ChatGPT or Claude, or the agents you build, can keep your business data and safely write to it. Think of it as Git for your business data. It works like this:

• Nothing is overwritten or deleted. A correction is a new entry. The original stays in the history.
• Every change has a name on it. Each agent gets its own key, so you can see which agent changed what, and when.
• Any change can be undone. Roll back everything one agent did in a window of time, with a preview first. A person can do it, or another agent can.
• The history can’t be quietly rewritten. If anyone edits or removes past entries, it shows.
• Retries are safe. An agent that retries after a timeout doesn’t create duplicates, and a write based on stale information is refused instead of clobbering newer work.

It’s worth being clear about one thing: AvianSuite isn’t a backup or an audit log bolted onto your other tools. It’s where the data lives. Your agents can still talk to Autotask, HubSpot or anything else through their APIs. But the business facts they create and rely on go into AvianSuite, where every version is kept.

It also won’t decide whether a fact is true. An agent can still write something wrong. The difference is that the wrong thing stays visible, you know which agent wrote it, and fixing it takes one step.

Watch the demo: an agent deletes records, a second agent spots it, and the data comes back.

Open source at the core

AvianSuite runs on Stellar Jay, an open-source store I’ve released under the AGPL. If you’re going to trust something with the data your business runs on, you should be able to read the code. You can also run it yourself for free.

AvianSuite is the hosted version: I run Stellar Jay for you, handle the servers, and give you readable tables and records on top.

Try it

The fastest way in is to let your agent connect itself. AvianSuite has an MCP server, so in Claude Code it’s one line:

claude mcp add --transport http aviansuite https://mcp.aviansuite.com/mcp

In Cursor, Claude or ChatGPT, add a connector with the same URL.

Plan	Price	What you get
Hosted AvianSuite	$20 / month per store	5 GB, 10 agent connections, 14-day free trial; $10 per extra 5 GB, $5 per 20 extra connections
Self-hosted Stellar Jay	Free (AGPL)	The full open-source core, on your own server

At $20 a month, it’s meant to go on every client project without a sales call. If you build agents for clients, that’s who I built it for first.

Back to that question

When a client asks “can it just do it?”, I can finally say yes without crossing my fingers. The agent does the work. If it gets something wrong, we can see exactly what it did and put things back.

I think every business that lets agents do work will need this. Give your AI agents write access. Keep every version of your business.

Start your free trial at aviansuite.com · Star Stellar Jay on GitHub

If you’re building agents for clients and want help setting this up, email me at kyle@futureperfect.work. I’d like to hear what you’re working on.