Sign in Subscribe

Can new companies be built as MCPs?

Can new companies be built as MCPs?

As I route more and more of my time and workflow through Cowork, I've found myself asking our software + saas providers to offer MCPs (if they haven't already). Their existing dashboards are observable interfaces, but between their MCPs and my Cowork harness, I now have an evolving programmable interface that I can customize to my workflow needs without relinquishing dependence on key software in our stack and without having to do my own custom data orchestration (which I know I would likely struggle with). 

Amidst concerns of foundation model creep, I am in realtime watching Cowork consume more and more of my workflow attention while also increasing my own reliance on our existing software stack (via MCP). 

For example, we use a product called Harmonic for enriched network and people data. I could, in theory, go build a bunch of scrapers on LinkedIn and cobble together the same information or even have Claude attempt to do this through a variety of scripts built with Claude code. But that would take time and we could just pay Harmonic to do the grunt work of acquiring, cleaning, and structuring that data. I rarely go to Harmonic’s dashboard but with their MCP, that enriched context flows directly into my Cowork harness and it's incredibly useful.

Similar story with Concrete, a portfolio management tool that screens our email for updates from portfolio companies, then packages it all into structured portfolio intelligence. Their dashboard offers value to observe insights across the portfolio but via MCP and Cowork, I could query for example “What customer intros is portfolio company X looking for and who in the broader BTV network can help facilitate?”. My Cowork harness would utilize Concrete, Harmonic and Gmail via MCP to surface intro requests across the portfolio, find pathways in the BTV network and draft emails to be sent all with one prompt.

Our dependency on software vendors is increasing not for their UI, but for their ability to package enriched, workflow-ready context that my Cowork harness can consume. This has me thinking about what kinds of companies and products can be built solely at the MCP layer.

Expanding on MCP companies

If a general harness such as Cowork is where we do more of our work, and the harness gets better as more personal, enriched context flows into it, then the companies providing and maintaining that context could occupy a genuinely valuable position. And buying vs building said context readiness is a reasonable tradeoff. 

I'm not suggesting that all application-layer opportunities collapse into the MCP-layer. There will absolutely be companies that need to go full-stack and build their own harnesses from the start. But I do wonder what an MCP-layer company could look like. A loose categorization might be Read-MCP and Write-MCP businesses?

Read-MCP businesses serve enriched context into the harness. Harmonic offers people and network data. Concrete offers portfolio intelligence. Another example is Granola that offers meeting summaries. Their relevance is in data quality, coverage, and freshness and their value is in how they acquire and stitch that data in the first place. Their moat might be in the breadth of data they are able to acquire or perhaps they have a better transformation + enrichment layer, or both? 

Write-MCP businesses execute actions and write into a common database. These could be systems of record (think ERPs), general ledgers, brokerages, payment networks, regulatory endpoints. Their value is in trust, authorization, and correctness. You need compliance + liability guarantees and permission structures to write into these systems.

The read/write distinction matters because the moat profiles are different. Read businesses defend on data acquisition. Write businesses defend on being trusted intermediaries to consensus-based formats, networks, or regulated systems.

Don't incumbents just add their own MCP?

The obvious objection: Salesforce, QuickBooks, Bloomberg all ship their MCPs. One counter to this is the innovator’s dilemma wherein their existing business models depend heavily on clicks and views on their own dashboards and observable surface areas. We saw this play out in Salesforce gating Slack data. 

All said, context is not zero sum. The volume of new data and therefore context will exponentially grow from here. So it's not only about exposing existing data and context but also about gleaning and stitching existing fragmented + unstructured data into enriched context (email correspondence with attached pdf purchase orders at a manufacturing business) and also becoming the primary conduit for all future context generated in this AI world (like Granola does with meeting notes). 

The new MCP-layer companies are rethinking where new forms of context exist and how it gets gathered in the first place. Granola is the clearest example. Instead of joining Zoom calls and becoming the awkward third wheel (like every other notetaker), it found a way to remain silent yet ambient while also capturing near-perfect context. Product innovation and a PLG motion gave Granola unfair and growing access to critical (and dynamic) company context. And with the Granola MCP, my Cowork experience accesses this context. Just because I don’t need to go to the Granola app, does not mean Granola is any less important in my system of work.

What are some context gathering wedges? 

Easier said than done but a few that come to mind are:

  • With zero to low friction, embed into existing workflow moments. Your customer does not necessarily require a new dashboard to monitor (unless the user wants to)
  • Transform and enrich context from one state to a dramatically better one and make that your core competency (buyers won't care enough to build this or it's too high of an effort to build, maintain and worry about)
  • Work with businesses to map, collect and create context graphs unique to them and “MCP-able” into general harnesses like Cowork. Kind of what Glean does for enterprises and I talk about a potential SMB/Midmarket approach here too. 

Tensions I’m thinking through

A few strategic tensions that emerge are:

  • Do the motions of company building demand that a Read-MCP business move into owning Write-MCP scope too, a la system of record? And is the MCP layer fat enough of an opportunity or do you need to build your own harness too?
  • Foundation model creep is still a major concern, and there are context surface areas they will of course lean heavily into. Growing companies have no choice but to expose their MCPs but what happens when the scope creep gets too close? Do they just hope that their own harness and Write-MCP scope is developed enough by then?

I'm still early in sharpening this and would genuinely love to hear from founders and builders thinking about this layer. What am I missing?