MCP support has quietly become a buying criterion

Update, 20 August 2026. A full vendor re-verification has overtaken parts of this piece, and rather than quietly edit it we are saying what changed. The index now covers thirty-three tools rather than eleven. The top score on this axis is Airtable at 4.8 and the bottom is Trickle at 2.5. Most importantly, a hosted MCP server is no longer a rarity worth remarking on: Xano, Retool, Softr, WeWeb, Airtable, FlutterFlow and Bravo Studio all publish one. What is still unusual is the price of programmable access. Totalum is the only tool on this roster with a published, flat, self-serve per-call price, USD 0.07 per credit, available at every tier including free. Everywhere else the answer is a plan gate, a quote or silence.

When we designed our scoring axes we gave API and MCP access five percent, on the reasoning that few buyers would exercise it. The community reviews arriving since launch suggest we may have called that slightly low.

The pattern in the feedback is consistent. Buyers who write any code at all want to drive the platform from outside it: seed data from a script, run a schema change in CI, let their own agent read and write the data model.

The spread is unusually wide

The spread on this axis runs from 4.8 to 2.5, which is one of the widest ranges of any axis we measure. On most axes the leaders and the laggards are separated by about a star. Here it is closer to two, and the gap is structural rather than a roadmap item.

The tools that score well tend to have been built by teams who use their own product programmatically. The tools that score badly tend to be the ones where the chat window is the only interface, by design.

Publishing one is now normal. Pricing it is not

Seven vendors in the index publish an MCP server: Xano, Retool, Softr, WeWeb, Airtable, FlutterFlow and Bravo Studio. Airtable has gone furthest, and now leads the axis at 4.8 on the strength of an official MCP server, one-click OAuth for Claude and ChatGPT, an npm CLI and a native agent runtime.

Read the gate, not the feature list. WeWeb's MCP is effectively a Pro-tier feature at USD 50 a month per developer. FlutterFlow's CLI, the piece that lets Claude Code or Codex drive it, is a Business-tier feature at USD 150 for the first seat. Softr publishes API and MCP access from its entry paid tier with the caveat that certain endpoints may be limited to higher plans. Retool's Platform API is Enterprise-only with no published per-call price.

Two absences are worth naming because they are often assumed the other way. Bubble has no official MCP server: its own founder AMA in March 2026 called it "on our radar, but not on the roadmap yet", the editor API is not open either, and everything circulating under that name is a community project. Draftbit's Pro tier adds custom MCP server integrations, which means Draftbit consuming your MCP servers, not publishing one of its own.

Why it matters more than five percent suggests

The argument our reviewers make is that a platform without programmable access cannot be part of a larger system. It can be a product but not a component. As more teams run their own agents against their own infrastructure, a builder that cannot be scripted becomes a dead end regardless of how good its generation quality is.

We are not changing the weighting mid-quarter, because a published weighting that moves whenever we get feedback is not a methodology. It is on the agenda for the next review of the axes, and if it changes we will publish the before and after.

What to test

Ask for the API documentation link before you buy. If the answer is a support ticket, you have learned something. Then check whether there is an MCP server, which plan it starts on, what a call costs, and whether schema changes are available through it or only through the interface.

---

Publisher's note, August 23, 2026. This piece used to carry a publication date from before App Builder Index was a live site. An audit of our archive found the stamp was applied when the index was seeded rather than when the article went up, so the date above has been corrected to the real one. The article itself is unchanged.