Websites that work with visiting agents.

A human visitor converts through persuasion. An agent is immune to persuasion; it converts through exchange. The Agent Surface encodes that exchange as WebMCP tools: every tool gives the agent a real capability and gives the site a real signal, by contract.

WebMCP 0 tools live on this page In production on 3 real sites MIT

Reading this in a WebMCP browser? Ask your agent:

What tools does this site offer? Call about_this_surface.

Works in ChatGPT's desktop browser today, or Chrome 149+ with chrome://flags/#enable-webmcp-testing enabled.

The exchange ledger

StageThe agent getsThe site gets
AnswerCanonical facts in one callIntent with declared purpose
FitAn answer computed for its buyerA self-qualified account, no form
WorkThe site does part of the agent's jobDeep engagement a human watches
AdvanceA computed offer, or a capability to keepA transaction, attributed

Three archetypes, three live demos

Product: the agent does the trial

Search a real 100M-company graph anonymously, build a target list in a workspace the human watches, then claim the work by connecting the vendor's MCP. Reference: graph8.com.

Service: the agent scopes the work

Market size, channel benchmarks with their basis, case-study proof: a feasibility brief instead of a discovery call. Reference: cience.com.

Authority: the agent grades the work

Score a team against a published standard; benchmarks return with citations preformatted. Reference: tenbound.com.

The honesty rule

Agents remember, retry, and tell their users which sources burned them. An Agent Surface answers "not a fit" when that is true, labels model assumptions as assumptions, quotes observed numbers as ranges with attribution, and marks the untested as untested. It costs one bad-fit lead and buys the successor to SEO ranking: agent trust.

Read the full spec (SPEC.md)