agent-preview/v1 — One Fetch, Every URL, Open Invitation

Mnemom now serves a structured agent-preview/v1 manifest at every prerendered URL on the site — a one-fetch, MIME-tagged JSON sibling that tells an agent everything it needs to know about a link before deciding what to do with it: title, summary, canonical, type taxonomy, representations (HTML, markdown mirror, OG image, JSON-LD), and per-link context like author, published date, tags, reputation grade, claim status.

If you've ever watched an agentic client try to read a webpage, you know what's broken. The agent fetches HTML — usually 300 KB of layout chrome it doesn't need. Or it fetches the markdown mirror, which has body text but no metadata. Or it fetches the OG image, which is a 1200×630 PNG. Three round-trips, zero of them structured.

agent-preview/v1 collapses that to one fetch. The shape is published at mnemom.ai/spec/agent-preview/v1. The same content framed as a proposed standard — Internet-Draft style, with schema, discovery, MIME registration, security considerations, and an explicit IANA path — is now live at mnemom.ai/spec/agent-preview/v1/rfc.

Why an RFC, why now

We're the reference implementation. We're not the gatekeeper. The schema is intentionally vendor-neutral: nothing in it depends on Mnemom, our API, or our brand. Any site can publish manifests, advertise discovery via <link rel="alternate" type="application/vnd.agent-preview+json" href="<url>.preview.json">, and become one-fetch addressable for agents — no permission needed.

We wrote the RFC because we want to be wrong about being the only ones doing this. The MIME type starts as a vendor-tree entry per RFC 6838. If three or four independent implementations ship — different stacks, different domains — we'll file for promotion to the Standards Tree, drop the vnd. prefix, and the registered MIME becomes the canonical one. The schema doesn't change as a function of registration status; adopters can ship today.

What's in the box

The substrate has been live since PR #412:

• scripts/build-preview-cards.mjs — the open-source build-time generator. Reads prerendered HTML, extracts metadata, writes the <url>.preview.json sibling.
• client/lib/preview-types.ts — the normative TypeScript types.
• Discovery via <link rel="alternate"> — every prerendered page's head carries it; the agent never has to guess.
• Nightly verification — the preview-surface commitment on /for-agents validates every manifest against the v1 schema; if anything drifts, the watchdog files an issue.

PR #414 wired four dynamic-content surfaces (blog, research, agent reputation, coherence reports) through the same construction path so their per-link context carries real data — author, reading time, tags, grade, score — instead of empty stubs.

Read it. Build it. Tell us.

If you ship structured metadata anywhere — a blog, a docs site, a profile directory — agent-preview/v1 works for you. The schema is forgiving: required fields are minimal, every context field is optional, unknown types are tolerated. Start with the implementer spec, check the RFC framing for the formal language, and the reference implementation is open at github.com/mnemom/mnemom-website — scripts/build-preview-cards.mjs is the only file you'd need to fork.

If you ship a preview surface on a domain you control, open an issue and we'll add you to the adopters list. The goal isn't ownership; it's that an agent fetching any URL in 2027 gets a structured answer in one round-trip — and nobody had to ask permission to make it true.