Access should follow the job. That goes for coworkers, clients, and the AI tools we now paste our tokens into without a second thought.
Select any sentence in this post to leave a margin note on it.
Did this land?
I'm pretty strict about access. Tokens, API keys, permissions, passwords, all of it.
My work laptop has production access, so I never leave it unlocked. Nobody uses it except me. Not even for a quick "can I just check something".
Even the client gets less access than me on some things. I trust them, that's not the point. I'm the one doing the work there, so I get the access the work needs. They don't need it, so they don't get it.
Some people find this a bit much. I think it's just normal. Give out as little as you can and know who has what.
So it's been kind of weird watching the last year.
People who would never share their password with a coworker are now pasting API keys into AI chats. Giving coding agents the prod database URL. Connecting random tools to their email, GitHub, cloud accounts, with whatever permissions the setup guide asked for.
MCP made it even easier. You find a server for whatever you want Claude to manage, paste one line in your config, put your token in it, and it works. Two minutes.
What most people don't think about is that an MCP server isn't "the AI". It's just a program someone wrote. It runs on your machine, it holds your token, and it does whatever its code says. And it keeps doing that after every update you never read.
If you think about doing the same with a person, it sounds stupid. Some guy you've never met offers to manage your server. Seems nice, profile looks fine. So you give him full admin and never check what he does with it.
Nobody would do that. But with a tool it somehow feels fine, because it's convenient and everyone else is doing it.
I hit this myself. I wanted Claude to manage my Discord server, The Basement, and Discord doesn't have an official MCP server. There are community ones, and honestly some of them look fine.
But all of them wanted my bot token. And that bot has real permissions on a server with real people in it.
So either I read some stranger's whole codebase, and then every update after that, or I write a small one myself. Reading would've taken longer. So I wrote my own.
I built it the same way I deal with access anywhere else. The first version was read-only. It could see everything and change nothing. The second version didn't add a single Discord feature, only safety stuff: a read-only mode, a list of servers the bot is allowed to touch, a way to turn off single tools, and a check that won't even start the server if a tool doesn't say whether it's destructive.
Only after that I added the write tools. And I tested those on a throwaway server where nobody lives, not on The Basement.
Probably sounds like a lot for a side project. It really wasn't. It's the same thing as locking my laptop, just in code.
I'm not saying everyone should build their own tools. Most people won't, and that's fine.
Just notice what you're giving away. Before you paste a token somewhere, take a few seconds. What can this key actually do? Does this tool need all of that or just read access? Who wrote the thing that's going to hold it?
Most platforms let you make a token with limited scope. Almost nobody does it, because the full access one is already right there.
Being strict costs a few seconds. Being loose costs you one bad day.