Hello fellow keepers of numbers,

This week, we saw some of the largest companies in the AI race partner together to release a new, interesting framework for agent plugins. Notably missing from the list of partners is Anthropic, who have also been suspiciously quiet of late outside of their Opus 5 release.

We also got another AI tax prep software that launched, coming out of the startup accelerator Y Combinator. Plus, stick around for a demo on how to create a plugin for client onboarding using the new agent plugins standard released by OpenAI.

The Latest

OpenAI introduced Agent Plugins, an open standard developed with Amazon, Cursor, GitHub, Microsoft, and Vercel that gives AI tools a shared format for reusable skills and software connections. Version 1.0.0 launches with support across ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code.

Today, a plugin built for one AI tool often has to be rearranged or rebuilt before another can use it. Agent Plugins creates one standard package that can include the instructions an agent follows, the outside tools it can access, and additional features made for a specific platform.

The standard covers how plugins are packaged and recognized by supported AI tools. Distribution, installation, permissions, and platform-specific behavior remain under each company's control.

Agent Plugins is openly licensed and developed in public. Its initial Technical Steering Committee includes maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel.

Why it’s important for us:

This framework is actually very simple, but could be incredibly important. So let's break it down for a second.

An Agent Plugin is essentially a folder. Inside of that folder there are a few things. It looks something like this:

my-plugin/
├── plugin.json
├── skills/
│   └── do-the-job/
│       └── SKILL.md
└── mcp.json

A plugin.json file identifies the plugin.

The skills folder contains the instructions for completing different tasks. It can contain as many skills as you'd like, and can include skills you already use today.

The mcp.json file contains a list of the MCP servers the plugin can use. MCP servers are how you add a connector in Claude or ChatGPT. This is the same, but added in a different way through the mcp.json file.

For an accounting firm, I'd think about a plugin as an entire process. For example, a monthly close.

The MCP file could connect the plugin to QuickBooks, Gusto, PayPal, and whatever else the firm uses. Then each step could be its own skill: reconcile cash, record the journal entries, review the financials, prep the reporting package.

Once that's built, you can just do "@monthly-close Run the July close." It understands a monthly close, applies the skills saved in the plugin, and uses the software tools via the MCPs.

None of those individual pieces are new. We already have skills. We already have MCPs. Plugins have existed for a while, and Anthropic already has a lot you can download in Claude.

The important part is portability. Right now, someone who uses both ChatGPT and Claude can’t build one plugin and carry it between the two tools unless they manipulate it to be specific to each. The packaging and setup are different. A developer sharing a plugin publicly may also have to maintain separate versions for different AI tools, even though each version is doing essentially the same work.

This standard attempts to fix that. There are a few open questions though.

But what happens when two plugins contain a skill with the same name? Client A and Client B might both have a reconcile-paypal skill, but with different instructions. If I ask the AI to use reconcile-paypal on its own, how does it know which one I mean?

Anthropic’s existing plugin framework already addresses this by adding the plugin name to the skill name. Client A’s version would be client-a-monthly-close:reconcile-paypal. I suspect this standard will need to adopt something similar.

The other hurdle is that Anthropic is not on the list of initial partners. OpenAI, Microsoft, and Cursor are all part of the committee. Anthropic is not.

Not sure if they were intentionally left out or chose not to participate. Either way, it's unfortunate. Hopefully, we can get everyone on board with the same framework at some point. It could be extremely useful when it comes to making AI useful for accounting work.

Luca IQ launched a professional tax platform for CPA firms after passing IRS Assurance Testing and becoming an Authorized IRS e-file Provider. The company says it is the first AI-native tax engine authorized by the IRS and the first new professional tax engine to clear the process in decades.

The platform reads W-2s, 1099s, K-1s, brokerage statements, and other source documents, then maps the extracted values to the corresponding lines of a tax return. AI handles the document reading and mapping, while a rules-based engine performs the tax calculations.

Luca IQ combines client intake, document collection, workpaper preparation, tax calculation, review, Form 8879 signatures, delivery, and electronic filing in one workflow. Every calculated field retains a source trail, and a tax professional reviews the return before filing.

The platform is currently available to accounting firms through a demo request. Luca IQ has not published pricing.

Why it’s important for us:

Yet another AI tax prep vendor in the market. I suppose it's good because more competition should drive down prices.

Based on the video, I'm not convinced Luca is doing anything materially different than what's already available in the market. It currently appears limited to 1040s, which is where most other AI tax prep vendors also focus.

There are a few interesting things to note though.

Luca is touting an API that allows you to build integrations and connect directly to the tax engine. The ability to connect it with Claude, ChatGPT, and others to build into existing workflows could be very useful.

They're also an approved e-filer with the IRS. This replaces the existing tax software. To my knowledge, the only other AI tax prep vendor doing this today is Instead.

So, definitely some positives.

Existing tax prep software sucks. It's old and doesn't play well with other software. I think nearly all of us can agree that we deserve something better.

But ironically, it's also one of the hardest things to convince tax accountants to change. Which means the change management that comes along with a firm transitioning to something like Luca is fairly painful.

And if you're preparing both individual and entity returns, you'd still need your existing tax prep software for 1065s and 1120s because it seems like Luca is only for 1040s right now. Ahh yes, one software for the price of two! What a deal.

I think Silicon Valley is figuring out what we've all known for quite some time. Filing accurate tax returns is a little harder than it might seem on the surface.

Trending News

Microsoft added a feature that lets Copilot in Excel analyze the underlying data from an attached Power BI report: Microsoft is making some interesting moves. Copilot itself is still weak compared to Opus 5, Sonnet 5, and GPT-5.6. But Microsoft is adding more capabilities and integrations with the rest of the Microsoft stack, so it's at least becoming useful.

Aider added client questions and AI document validation to its period-close product for U.S. customers: Aider updates at this point are also Karbon updates. This is a nice addition to the tool. One more tool in the toolkit when it comes to automating manual tasks.

Cursor launched Google Workspace plugins that let its agents work across Gmail, Drive, Calendar, Docs, and Sheets: This is just another step in the right direction to become a true competitor to Claude Cowork and the new ChatGPT app.

OpenAI added an Activity view that collects pending conversations and recent project updates in the ChatGPT desktop app: The ChatGPT app is doing so many great things lately. This is another great update. It's a cool view to easily see the things you're working on and unread threads.

Aprio earned accreditation to certify organizations against the ISO 42001 AI governance standard: Aprio can now review whether a company has proper guardrails around how it uses AI and give it an official stamp of approval. It's probably a needed service line, but things are still moving so quickly that the goalposts move constantly with AI.

Bending Spoons agreed to acquire Airtable for $1.285 billion, about 80% below its 2021 valuation: This company has a history of gutting businesses they acquire and trying to drive up profits. Lots of firms use Airtable, so this is one to be wary of moving forward.

Alibaba released Qwen3.8-Max and promised open weights for its largest model: Here is the obligatory reminder that Chinese labs keep shipping frontier-ish models for pennies. I'm still skeptical that it's as good as the top-end models, but it's close. And for accounting, we don't always need the top-tier models. It's important to at least pay attention to this.

OpenAI revealed Astra after the model solved ten long-open math problems with machine-verified proofs: Square root. PEMDAS. Integrals. Tangents. Math stuff, yeah. As a math nerd myself, this is cool because we're somewhat limited scientifically with what we currently understand in mathematics. AI could unlock some very important scientific advancements for us.

Workday made its Financial Audit Agent generally available, claiming a 75% reduction in evidence-package preparation time: Agents are everywhere now. It's honestly impressive if you go to a website and don't see a mention of AI agents. I guess it's cool if it's actually useful for audit.

Put It to Work

I decided to put the new OpenAI agent plugins standard to the test. Like I mentioned above, the idea isn’t all that different from how plugins have been created previously. But even in the ChatGPT app itself, they’ve got some work to do to make the plugins easy to use and useful.

Weekly Random

Passwords are dumb. They're so dumb that a lot of us pay for a software to store all of them.

Think about how ridiculous this system is. We're supposed to use a different, random, lengthy combination of numbers, letters, and special characters for every website where we login. Sometimes they get compromised without us knowing, and when we find out, we have to rotate to another different, lengthy, random combination.

So you create passwords you'll never remember. Maybe you never even looked at them in the first place if your password manager creates them for you.

Your password manager hopefully works everywhere, but what if some of your passwords are in Apple Passwords?

And of course, the password manager also needs a password. And if you forget that one... pretty sure you have to cancel every card linked to your subscriptions and start over.

I saw Google is rolling out a new option that offers another way to sign in: selfie video. You record a video of yourself ahead of time. Signing in with video later just requires you to do another selfie video to get access.

Selfie verification is already used for more serious stuff too. ID.me uses it to verify people accessing IRS and other government services. And Proof compare a selfie to your government ID before some signatures and notarizations.

This got me thinking. Why the hell do we have passwords anymore? Can't we just do this? Apple has been doing it for like a decade with Face ID.

I genuinely have no clue how secure it is. But if a criminal can perfectly recreate my face and head movements to hack into my Taylor Swift fangirl burner Gmail account, then they can probably hack into my password manager to steal every single password I've ever used.