What we found.
Where our agent got to call a surface, the call worked at 94% or more of those products. What stopped it was getting to the call.
What the 99 products offer
| Surface | Products | Share |
|---|---|---|
| API | 97 | 98% |
| MCP server | 87 | 88% |
| CLI | 59 | 60% |
| SDK | 51 | 52% |
| Agent payment | 13 | 13% |
Share that worked, by surface
| Surface | When called | When offered |
|---|---|---|
| API | 100% (62 of 62) | 64% (62 of 97) |
| MCP server | 94% (48 of 51) | 55% (48 of 87) |
| CLI | 98% (48 of 49) | 81% (48 of 59) |
| SDK | 96% (25 of 26) | 49% (25 of 51) |
| Agent payment | not counted | 23% (3 of 13) |
Agent payment: we don't count calls. It worked when a payment went through or a valid payment request came back.
- 48 of 59A CLI worked most often: at 48 of the 59 products that offer one.
- 25 of 51An SDK worked least often: our agent called one at 26 of the 51 products that offer one, and a call worked at 25.
- 3 of 13Agent payment: at 3 of the 13 products that offer one, a payment went through or the product answered with a valid payment request.
What to do.
For product owners
- List every API operation, MCP tool and CLI command on a page an agent can read without signing in.
- Give each surface one call that works on a free account, so an agent can prove it works before it builds on it.
- If you take agent payments, answer an unpaid call with a valid payment request that names the price.
For agent builders
- Start with the surface that worked most often, and keep a second one ready.
- Treat a surface that is only documented as untested: make one cheap call first.
- Before you build on an SDK or a CLI, check that it installs where your agent runs.
Method and limits.
How we tested
- Drives from 1 Oct 2026 to 1 Nov 2026, 201 products. 99 of them had their agent surfaces mapped from the docs.
- Offered: the docs list at least one API operation, MCP tool, CLI command or SDK method, or the drive found an agent-native way to pay (x402, MPP and the like).
- Called: our agent called at least one documented item of that surface. Worked: at least one of those calls came back done.
- Agent payment worked: a payment went through, or the product answered with a valid payment request for its core work.
Limits
- Each product as it was on the day of its newest drive.
- A drive calls a sample of each surface, so a surface that didn't work for our agent may still work for some calls.
- A surface fewer than 5 products offer has no row.
Cite this
AgentDrives, “Which agent surfaces products offer, and how many worked in our drives”, 1 November 2026, https:// agentdrives.com/ findings/ agent- surfaces.
@misc{agentdrives_agent_surfaces_2026,
author = {{AgentDrives}},
title = {Which agent surfaces products offer, and how many worked in our drives},
year = {2026},
month = nov,
url = {https://agentdrives.com/findings/agent-surfaces},
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.