Skip to main content
wr.fi

WRFI

The open standard for handing off AI work

Handoff — protocol 1.2.0, live & probe-tested · Discovery — v1.0 draft · spec (CC-BY-4.0)

A WRFI handoff is a durable, versioned object that lets unfinished AI work wait for its next participant — another session, another vendor's tool, or a human — without re-explaining everything. WRFI transports work-in-progress, not merely content. The standard has two layers and a deliberately tiny core.

Two layers

WRFI Handoff

The portable continuation object and its lifecycle: what a handoff contains, how it is versioned, and how participants coordinate around it safely. This is the layer that makes work continuable.

WRFI Discovery

How an AI-capable web page declares machine-readable actions — endpoints, schemas, auth, risks — so an agent with nothing but web access can find and use them. This is the layer that makes the handoff reachable with zero configuration.

Discovery teaches an agent how to reach a service; Handoff defines what it exchanges once there. A conforming implementation can adopt either layer or both.

The core is small on purpose

A useful independent implementation of the Handoff core should be buildable in a weekend. It requires only:

  • a stable URL identity for the handoff object;
  • the content itself;
  • the task layer — objective, requested action, status, handoff message;
  • a monotonic version;
  • a machine-readable handoff view (context, history, update paths in one fetch);
  • update preconditions — no silent overwrites, conflicts fail loudly.

Everything else is an extension, not an obligation: append streams, review/acceptance loops, workspace manifests (portable environments), provenance metadata, federation, and claims. wr.fi implements the core plus these extensions; a conforming implementation does not have to.

Conformance, not ownership

A WRFI implementation isn't compatible because it says so — it runs the conformance suite. Every guarantee the standard makes maps to an executable probe, and wr.fi/conformance shows the live results against this deployment, re-verified continuously. The suite is open and runnable against any implementation: matrix-probe.mjs <your-url>.

Standard, service, implementations

WRFIThe neutral specification (CC-BY-4.0). Small core, clean extensions.
wr.fiThe hosted reference service, run by Kurikkai Oy.
wrfi-cli / wrfi-mcpReference clients (MIT), at github.com/wrfi.
wrifyThe server reference implementation (AGPL-3.0) — designed for self-hosting and federation; public source release in progress.

Compatibility is determined by conformance, not by who runs the server. The “wr.fi”/“WRFI” marks are reserved; the standard is not.

Discovery in practice

Honest assessment: no single instruction method reliably reaches every AI tool today, so Discovery is specified as layered redundancy — the agent hits whichever layer it understands first.

MethodWorks?How an agent finds it
/.well-known/wrfiYesCanonical bootstrap: protocol version + surface map
/llms.txtYes<link rel="alternate"> or direct URL
Visible page contentYesAgent reads the page as text — most reliable layer
Self-documenting APIYesGET on the endpoint returns schema + examples
Content negotiationYesNon-browser user agents get JSON/text
OpenAPISometimesSome tools fetch specs
<meta name="ai:instructions">EmergingIn <head>, always server-rendered
<script type="text/wrfi-instructions">FutureStructured block for native tooling support

wr.fi serves all of these simultaneously from one source of truth. Full layer definitions and the script-tag format are in the Discovery spec.

WRFI vs OpenAPI vs MCP

WRFIOpenAPIMCP
CarriesWork-in-progressAPI shapeLive tools
SetupZeroSpec fileServer + config
Both sides online?Never requiredN/AYes
Best forContinuing unfinished workFull APIsAgent tools

They complement each other — a handoff can even declare the MCP servers its work needs.

WRFI is an open standard (CC-BY-4.0), served canonically from this site: Handoff protocol · Discovery layer · Live conformance · hello@wr.fi