In a recent Claude Code session I tried using the Google Docs, Drive, and Sheets MCP and was honestly surprised at how limited it felt. It was hard to get anything meaningful done because it just did not expose enough capability to be useful in practice. In hindsight, that frustration was probably a good thing. I ended up skipping MCP entirely and using the LaTeX API plus its plugin ecosystem, and the result was far beyond anything I could have realistically produced through Docs anyway.
I have seen a similar pattern with Canva’s MCP. I pay for Pro, but the one feature that would actually make MCP useful, Auto Fill, is gated behind an enterprise plan. So the surface is there, but the real power is locked away.
I get that this is still the wild west for MCP, and I agree with the OP’s general take. But right now there is a big gap between "integration exists" and "integration is actually useful." Personally, I am more excited about where something like WebMCP could go, where the default assumption is full capability rather than a restricted subset.
The result is less that I want to go to Figma directly and more that I just want to skip it entirely. So, assuming the power of these aggregator agents keeps growing, the onus is on these tools to create useful integrations or get subsumed by a model capability or another tool with a better integration. It sounds like your experience is similar - you bypassed the tools with bad integrations instead of going to them directly.
You have to be very very careful with this stuff. These SaaS companies have tons of paying customers giving them thousands of dollars a month. If customers mess up with an officially supported MCP and delete their assets or break implementations or DDOS their own site it’d be nightmare for sales / support.
It makes sense they very slowly transitioned from read-only to limited write. You have to carefully beta test. Both figuring out the guardrails and finding usecases where it actually works well. The only way to do that (properly) is a slow drip release cycle.
I'm sure that's part of it but I think it's a very small part of the story. They've been pushing hard on internal AI creation tools for a while, and those clearly didn't take hold.
IMHO MCP is the AppleScript/Shortcuts part of the internet. Never really took on Apple platforms (except for some hardcore fans); not sure it will ever actually take on the internet…
I had same experience with Databricks. The built-in MCP offering is very limited (querying, etc.) but there are community-built projects that offer the full scope (creating ETL jobs, etc.). I would prefer not to have to go through the hassle of getting some random project on GitHub added as an artifact and deal with the updates.
All SaaS-built MCP servers should cover the entirety their existing API functionality. I know it sounds like a lot but I really don't think it is an unreasonable expectation.
I have seen a similar pattern with Canva’s MCP. I pay for Pro, but the one feature that would actually make MCP useful, Auto Fill, is gated behind an enterprise plan. So the surface is there, but the real power is locked away.
I get that this is still the wild west for MCP, and I agree with the OP’s general take. But right now there is a big gap between "integration exists" and "integration is actually useful." Personally, I am more excited about where something like WebMCP could go, where the default assumption is full capability rather than a restricted subset.