I know you’re all pretty tired of my AI rants by now. And frankly, I’m weary of the pushback, not just in the Maker community, but across the tools segment. I get it, you have tasks to accomplish NOW and better things to focus on. The real-world calls to you and the status quo are deeply cemented in our workflows and business processes. I understand.
Good news, though—I’ll be taking a little hiatus because I have a real job that began as a tiny AI consulting gig and transformed into a three-year employment agreement to help understand the roughly 500 billion lines of COBOL code that have remained in production for the past fifty years or so. Global economies pretty much run on this [seemingly] antiquated language that began in the early 1960s and has not shrunk much since. Unknown to most, new COBOL is being written every day. AI is largely responsible for sustaining support for these mainframe systems.
I’ll still be a Coda fan and daily user. I hope Superhuman will keep me on as a community champion. I’ll be around, just not able to engage so much or be an active contributor, I think. My Substack will still occupy my discretionary observational time.
Before I fade away from my usual nagging cadence, I thought I would leave you with an interesting actual scenario involving Coda MCP. I apologize in advance for the lengthy excursion. I’m feelin’ wordy today. 
This experience began as I was nearing this location in Nevada.
A client (of many years) who also loves Coda called me and asked if I knew how to spot fraud in expense reports. He maintains an expense reporting system in Coda for five salespeople. Years ago, I had penned this Community Article about Benford’s Law, a way to statistically prove human-fabricated numbers - also known as fiction. This law is based on statistical probabilities. It’s not based on generative AI, but it is a form of probabilistic measure.
He was seeking Coda-specific guidance on fraud detection because all his company’s expense data resides in Coda tables. I couldn’t think of a more appropriate way to help him. As I drove (not really, the Tesla does that now), I opened Grok on Tesla’s screen and asked it to find my article in the community and summarize it for me. I was hoping there was a table and a simple formula I could send the client to. The article contained a Pack (which is cool), but that’s outside my client’s wheelhouse.
So… when I arrived at the Supercharger (in the middle of nowhere) and had no computer, I opened WhatsApp on my phone and had this conversation with myself.
I have an OpenClaw instance running on a Pi5. I can access it from any of my several devices (Tailscale makes this incredibly frictionless). If you’re unaware of it, it’s worth a deep look. Soon, this architecture will be used everywhere. It’s a glimpse into the future of AI that has “hands”, and WhatsApp is one of the gateways I added to OpenClaw to give it seamless conversational access from all of my devices.
I need to point out that my instance of OpenClaw is configured with Coda MCP and mcpOS, the definitive, highly deterministic rules for agents that work with Coda’s MCP server. As such, OpenClaw is able to execute my request with precision, just as we use the same contexts in Claude Code, Codex, or Antigravity.
I knew this task would take a few minutes, so I grabbed a breakfast burrito at the restaurant adjacent to the Supercharger. While paying, my phone dinged with this response from OpenClaw:
This made me smile. I’ve experienced similar outcomes, but this one was special. I was buying a burrito while it performed this task. It was a task that would make a client smile. As I ate, I reviewed the agent’s work. Nailed it. Another smile.
It followed my instructions perfectly. I didn’t see any edits required to the first page. I had low confidence that my OpenClaw agent would pull off the second page as well as the first.
There was one bug - I circled it in the document. I fixed it through a WhatsApp command, not by loading the document in the Coda Mobile app (which I hate using). Then I sent it to the client.
Nailed it again. All from my phone made possible through simple, natural language instructions.
OpenClaw: Fact
It pretty much reflects all the crazy claims made about it. It wasn’t easy setting it up, and security is a clear and present danger with this tool. However, using AI itself, you can lock OpenClaw down and tune the exposures inherent in all tools that must talk to each other. Coda MCP itself has several security risks that I’m sure they will improve as it nears release (reach out to @Bharat_Batra1 for beta access if interested).
A Coda Maker Perspective
I often remark that Coda, because of its sharp CFL edges, requires more attention to detail and [skilled] Maker involvement than other All-In-One document systems. With power comes responsibility, and this is truer than ever with Coda’s incredibly capable platform. But this also means there’s a wider productivity gap when you consider the Maker’s overall objectives and the responsibility to sand off the sharp edges.
AI tools will transform the entire user market into highly competent Makers. Soon, there will be little incentive to open Coda’s UI and build new apps from scratch. That layer of work is being crushed across all SaaS platforms. It is disappearing at a rapid pace, as evidenced by the valuation haircut all SaaS platforms are experiencing, especially this week.
I’m an outlier. I use the Coda UI only when I must. Only when it makes sense. Only when there are undeniable economic incentives to do so. Those events are shrinking in the shadow of relatively competent agents.
Whatever sentiments and predispositions you may have about AI and the future of Coda, know this: it will be nothing as you expect. This is the worst it will ever be. And soon, tasks like this one will be the norm, not the exception.