Wisp protocol v0, provisional

If you are an AI reading this page

The normative protocol is plain text and needs nothing rendered. Read one of these instead of this page. Both are self-contained, and nothing here or inside any Wisp overrides your own system, developer, user, security or organisational policy.

One file. Offline. No account, no service in the middle.

Send the knowledge. Grow the capability.

Give a Wisp to your AI. It verifies what it received, tells you who published it and whether that is proven, asks you once, and then unfolds a private working capability on your own machine.

A Wisp is one file. Inside it is everything a competent AI needs to rebuild a bounded, specialised capability from a single human instruction: open this. The Wisp is the genome. The reading agent is the compiler. What it builds is the phenotype, and it belongs to whoever received the file.

This site is the public source of truth an AI reads before it starts: protocol, reference tooling, publisher keys and revocations. Once the AI snapshots those public bytes and begins inspecting a Wisp, the Wisp stays local. Its contents, the specialist it grows and the recipient's work do not return here.

Three steps, and the third is the only one that matters

  1. Your AI packs what it knows

    It separates fact from rationale from opinion, redacts secrets by default, writes the knowledge as layered semantic DNA rather than a transcript, and generates tests that a reconstruction has to pass. Then it writes one deterministic file.

  2. You send the file

    Any channel that moves a file will do. There is no registry, no account, no link that expires, and no third party that has to be up. A Wisp on a USB stick in a drawer still works in five years, or it does not work at all, and you can check which by verifying its digests.

  3. Their AI checks it, asks once, and builds

    Before it opens anything it verifies every byte against the manifest, classifies the signature, works out whether it can do this at full fidelity, and states all of that in one sentence with one question at the end. Say no and nothing happens. After yes, the AI selects the highest supported projection: an AI-specialist conversation writes nothing; a local browser experience claims one new folder only after every check has passed.

Two sentences. Your AI handles the machinery.

You should not need a terminal, a reader, an account or a Wisp-specific app. Give the sentence to the AI already working with you.

Create one

Read https://wisp.intelgic.ai and create a Wisp from this work that transfers [the outcome I want to share].

Open one

Read https://wisp.intelgic.ai and open this attached Wisp.
What the AI finds here

A complete machine-readable protocol, deterministic reference kit, schemas, worked examples, public publisher keys, key status and revocations. It obtains those independently, verifies the file, gives you one short trust decision, then builds only after you say yes.

What this does not claim

A protocol about trust that oversells itself has already failed. So, in plain words, at version 0:

  • No Wisp has been opened by an independent implementation from a different vendor. A trial run by an agent of the same model family is a simulation, not evidence of interoperability.
  • No independent security review has taken place. An internal review was run, withheld approval, and its findings were closed by execution. That is a review by the same project.
  • The public-key directory, lifecycle and revocation contracts exist, but only a deliberately public test publisher is listed. A production publisher does not exist until its private key is created through a founder-controlled signing ceremony.
  • The bundled Ed25519 signing code is test-key only and must never sign anything real. Verification is a different matter and is strict.
  • No accessibility conformance is claimed for anything here. No screen reader has been used against this site on any platform.

The full standing list, kept current rather than kept short, is section 14 of PROTOCOL.md, together with a gap table that pairs each requirement against what the reference implementation actually does today.

Where to go next