Braden James

curious about things | *not financial advice*

[Braden James @bradenjms]

@BillAckman @AnthropicAI Does that fall under act of God/ force majeure?

[Braden James @bradenjms]

This diesel export ban is killing my vibe. Don't buy crak kids

[Braden James @bradenjms]

@irastech Dang.. that’s painful. Gotta pay extra for those snapshots!

[Braden James @bradenjms]

Enterprise model for distributed skills

Useful when skills serve as SOPs but still need to be siloed

1.) dedicated server (cheap droplet is fine) to serve an MCP.

2.) MCP is just a pointer to a custom directory holding shared skills and shared databases.

3.) The MCP manager sets permissions on which roles get access to certain skills and data.

4.) Each user is connected and given a role.

The MCP can also be used to make custom applications that serve as connectors for all of the common apps that will work much better than the lazy off-the-shelf connectors they provide. Becomes one simple connector and all the functionality you could ever want.

Not to mention an "idle conversation" pull that you can use to keep track of employee ai usage and engagement if you're feeling "big brother-ish".

[Braden James @bradenjms]

That’s it! Get as much out of the sandbox as possible. BUT that’s the value of the idle pull. Sometimes in chat employees can come up with promising updates to an SOP. A idle conversation push programmed into the mcp allows you to analyze what the employee is iterating on in the sandbox and apply it company wide if it makes sense.

[Braden James @bradenjms]

That’s true. And actually beneficial if you have the machine authenticated on the git account housing the shared folder. Then it’s still one connector and the mcp points clients to the repo for pulling skills. Meanwhile machine can also be authenticated on several other platforms (think like QBO) where mcp will pull the skill to “create a balance sheet” from git and then pull data from qbo, run scripts on the machine, and push result to the user. Much easier for say a bookkeeper to ask for a balance sheet on chat and all of that happens in the background. Routed through the machine you own and control not the llm sandbox

[Braden James @bradenjms]

Wow… That’s a good question. Didn’t know that tweet would reach someone who got it.

Roles are actually too constrained and not all I use in practice. They are like a “starting point”. On machine you can spin up a UI and connect it to a user db. Each new user gets their role base but you can have checkboxes for skills and dbs that they can access through the mcp that may (or may not) be included in the role. Protocol reads the db for items they can or cannot access. Ideally you get them fully customized before first connection.

Revocation is for sure the hard part. You don’t want to change all the time because when changing you must force a reconnect or else the user gets different access based on conversation history (old convos get the old access and new convos get the new). Good news is you can hack the connection so that it’s a simple Oauth with a custom Google or Microsoft app set up on the machine.

Regarding auditable trail - each mcp has custom link for the user and every action is logged. Either through idle conversation history or through true log on the machine if they take action like updating scripts or databases. Can also enforce though the mcp a local git protocol so all changes by a user on machine can be rolled back. Writing to the machine can be super powerful because it gets chat out of its sandbox and in a domain you control. But those changes do need to be logged carefully for non-technical users

[Braden James @bradenjms]

@conneryoungBC Looks like the typical using MCP as a way to serve application products. This would be one of the lazy connectors I mentioned. Could be wrong though. I have no idea who those people are lol

[Braden James @bradenjms]

Tips:
The key is staying organized, doing a lot of tests, and managing md files really well. Also you just have to be brave and send it. In most case, if it messes something up, you'll live to fight again.

I want to end this with a note that I'm not an expert in any way. I hardly know what I'm doing.

But the proof is there. These systems have been tested real-time and they work. A month ago I had no idea all this time would be opened up to me. I now get to spend time researching what I love instead of managing headaches. It feels like freedom and all it took was 30 days of dedicated work.

Every business is different. Find what annoys you and get AI to make it easier.

That's all. If you made it this far hope you enjoyed the read!

[Braden James @bradenjms]

30 days ago I sat down with a mission to make AI useful to me.

It worked.

I'm going to share everything I did in 30 days and how it has completely changed my business.

*Note my agent's name is Bingo - named by my daughter*

[Braden James @bradenjms]

Infrastructure:
Bingo is a 'Frankenstein' of the openclaw infrastructure. I never liked openclaw proper, but it had some incredibly useful tips and ideas to run with. So I've molded it to fit my particular needs.

I have subscriptions for claude and codex (use both desktop apps locally and cli on the vm - and both of them think they are Bingo). They are!

I run everything on a vm (Bingo's home base). And yes, he is accessible "on-the-go".

[Braden James @bradenjms]

This company is Carmax, it was removed from SP500 on Oct 31. Every index fund forced to sell.

Obvious play and it worked out. Stock down ~20% since this tweet. Closed position because the thesis played out.

On to the next one!

[Braden James @bradenjms]

There is one company in the SP500 right now with a ~$6B market cap. The last 10 companies removed from the index were at ~$6B on date on of removal, only two show a positive return since.

Going short because I like those odds.

That’s inductive investing

[Tyrael @0xTyrael]

At the end of the Coinbase earnings call today, Brian Armstrong pulled up the mention market and rattled off all the words that weren't said yet

What a wild time to be alive