MCP or API: how much of a product can an AI agent run through each?

At 108 of 123 products with both an API and an MCP server, an AI agent can reach fewer of the product's features through MCP than through the API.

What an AI agent reaches and finishes, through the API and through MCP. Features the API reaches: 2,346 of 3,955, Features the MCP server reaches: 1,232 of 3,955, Finished through the API: 831 of 2,054, Finished through MCP: 173 of 297.

Drives 1 Oct 2026 to 1 Nov 2026123 productsStep UseSubject API coverageMarkdown

What we found.

The features each route reaches
Route	Features it reaches	Share
API	2,346 of 3,955	59%
MCP server	1,232 of 3,955	31%

What came of the features our agent ran through each route
Route	Features run	Finished	Done in part	Needed a paid plan	Stopped
API	2,054	831	77	156	990
MCP server	297	173	18	36	70

13 of the 123 reach more features through MCP than through the API.
208 of their 3,955 features are reachable only through MCP.
1,401 of their 3,955 features have neither an API endpoint nor an MCP tool.
Our agent finished 40% of the features it ran through the API and 58% of those it ran through MCP.
Through MCP, 12% of the features it ran needed a paid plan. Through the API, 8% did.
It finished at least one feature through MCP at 24 of the 50 products where it used MCP, and through the API at 68 of the 148 where it used the API.

What to do.

For product owners

List on your MCP server's docs page which features it covers and which need the API.
Give the MCP server the work agents come for: create, read back, change and delete the product's main objects.
Let the key an agent gets for the API work for the MCP server too.

For agent builders

Check the API when the MCP server lacks a feature.
Read the plan rules before you call a tool. A tool can be listed and still need a paid plan.

Method and limits.

How we tested

Drives from 1 Oct 2026 to 1 Nov 2026, 201 products. The docs of 123 of them name an API endpoint for at least one feature and an MCP tool for at least one.
For each product we list its features from its docs, each with the API endpoint and the MCP tool the docs name for it.
Our agent ran each feature through one route. A feature counts as finished when our agent read the result back and it matched.
The results by route count every product we drove where our agent ran a feature that way: 148 through the API, 50 through MCP.

Limits

Each product's docs and MCP server as they were on the day of its drive.
Each feature ran through one route, so the two results count different features.
A product our agent didn't get into has no results by route.
Counts under 5 are left out.

Sources

Counted 1 Nov 2026.

Cite this

AgentDrives, “MCP or API: how much of a product can an AI agent run through each?”, 1 November 2026, https://agentdrives.com/findings/mcp-vs-api.

@misc{agentdrives_mcp_vs_api_2026,
  author = {{AgentDrives}},
  title = {MCP or API: how much of a product can an AI agent run through each?},
  year = {2026},
  month = nov,
  url = {https://agentdrives.com/findings/mcp-vs-api},
  note = {Agent test drive, CC BY 4.0},
}

CC BY 4.0, with credit

See what an agent sees in yours.

Free: what an agent finds, any wall in its way and how many things to fix. The Audit ($499) gives you every fix.