Capability

A Capability is a declared, invocable action an origin can perform or vouch for. It is the origin's offer to do work, not merely to display content.

Most of the web is content. Things to be read. That is the right shape for a human with time and a screen. It is the wrong shape for an agent acting on someone's behalf, which does not want a paragraph to interpret but an action it can take. Capability over content is the shift from here is what we know to here is what you can invoke.

A capability is specified, not implied. It has a name, a description and defined inputs and outputs. It carries provenance. That specification is itself projected. Into JSON for an API, into llms.txt for discovery, into an MCP manifest for an agent, so a machine can find and use it without a bespoke integration or a human reading the docs. Discoverability is part of the contract, not an afterthought.

The first capabilities are deliberately modest and honest. Ask a question of the corpus and get a cited answer. Retrieve a definition and its relationships. Return the provenance of a claim. Expose research for retrieval. Each is a promise to do a small, verifiable thing well.

A promise is only as good as its backing. A capability that returns unattributed or unverifiable results is available but not trustworthy, and untrustworthy capabilities are invoked once and then routed around. So capability and provenance are two halves of the same idea. The offer to do work, and the evidence that the work can be believed.

Content tells you what an origin knows. Capability tells you what it can do for you. The second is the one that gets used.