Hardwire 0.7.0: three actions, one experiment

Hardwire 0.7.0

Agentic actions. Without the clicks.

Three actions from the visual demo are now executable local experiments: sign in,
invoice payment and settings. Change an input, see validation block it, approve the
exact local action and inspect the state that changed. The experiment uses the
same artwork and cinematic visual language as the film.

Try it

Open the experiment.
No app, account or API key is required.

With Node 22 or newer:

npx --yes --ignore-scripts --package=https://github.com/Automatron-workflows/hardwire/releases/download/v0.7.0/hardwire-0.7.0.tgz hardwire-experiment

The package comes from this GitHub release, not the unowned npm name hardwire.
The attached SHA256SUMS records its checksum. The release workflow verifies the
uploaded artifact by downloading and installing it in a fresh temporary project.

Included

1. A cinematic browser experiment with explicit local consent and invoice approval.
2. The hardwire-experiment command covering all three fixtures and invalid requests.
3. Shared artwork, readable errors, reset and a keyboard accessible inspector.
4. The existing five inspection tools over local stdio MCP, portable core and CLI.
5. Community guidance and approachable contribution tasks on GitHub.

Scope

Experimental release. Google sign in is a fixed fixture, not OAuth. Invoice
payment changes local state, not a balance. Local approval levels are assertions,
not verified identities. Compiled tools are descriptions; the toolkit MCP server
does not execute supplied actions. The film is not a speed benchmark.

npm and MCP Registry publication require a verified Automatron owned npm scope.
They are not part of this GitHub release.

Hardwire 0.6.0: try the toolkit in your browser

Hardwire 0.6.0: try the toolkit in your browser

Hardwire helps developers describe application actions and check their inputs,
declared effects and approval requirements before connecting them to an AI agent.

Open the playground.
No app, account, installation or API key is needed.

Try it

1. Search for a lamp in the sample shop.
2. Let Hardwire turn the form into an action contract.
3. Call the fixed demo search directly. Try the wrong input to see it rejected.
4. Download the inferred contract.

The demo uses the same inference and validation code as the CLI and MCP server.
Inputs remain in the tab. The only executable handler searches fixed fake data;
it cannot call arbitrary endpoints or perform writes.

What changed

Version 0.6.0 adds the public playground, a portable hardwire/browser entry point,
async Web Crypto hashing, a reproducible static build and GitHub Pages deployment.
The synchronous Node API, panel format and trace format remain compatible.
The local demo also serves the playground at /playground/.

Read the architecture
for module responsibilities, runtime boundaries and deployment details.

Install the versioned source

Use Node 22 or newer:

git clone --branch v0.6.0 --depth 1 https://github.com/Automatron-workflows/hardwire.git
cd hardwire
npm ci --ignore-scripts
npm run demo

Open http://127.0.0.1:8787/playground/. To connect an MCP client, run
npm run --silent mcp:config and follow the setup guide.

The release also includes hardwire-0.6.0.tgz and SHA256SUMS. Download both,
check the checksum and install the local tarball:

shasum -a 256 -c SHA256SUMS
npm install ./hardwire-0.6.0.tgz

MIT licensed. Hardwire is not published on npm and its package identity has not
been reserved. Do not use npm install hardwire to obtain this project.

Scope

This remains an experimental toolkit. It validates descriptions and arguments;
it does not log in, execute submitted ports or prove that an endpoint behaves
safely. The benchmark still contains three synthetic fixtures. General execution
and real website inference accuracy remain open work.

Hardwire 0.5.0: make software understandable to AI agents

Hardwire 0.5.0: make software understandable to AI agents

Hardwire is an open source toolkit that helps developers describe application
actions and check their inputs, declared effects and approval requirements
before connecting those actions to an AI agent.

Why

An agent integration needs more than a button label. Developers need a clear
description of what an action expects, what it changes and which approvals
apply. Hardwire makes those declarations explicit and provides tools to check
them before connecting them to application code.

What

This first public release includes a Node.js library, command line tools and a
local MCP server. Its five tools infer draft action descriptions from supplied
HTML, inspect contracts, validate arguments, generate MCP tool descriptions and
check trace consistency.

It is for developers and experimenters building agent integrations. You do not
need an app, hosted service or API key to try Hardwire. The code is MIT licensed.

How

Supply saved HTML or a JSON action description, review the draft and its declared
rules, then validate example arguments. You get a contract you can inspect and
test. Connecting it to a real service still requires a trusted application handler
with authorization and approval enforcement.

Use Node 22+ (24 recommended). To try this exact release:

git clone --branch v0.5.0 --depth 1 https://github.com/Automatron-workflows/hardwire.git
cd hardwire
npm ci --ignore-scripts
npm run test:mcp
npm run --silent mcp:config

Merge the generated entry into your local MCP client's configuration. Follow
the setup guide
for a first prompt and troubleshooting. No application ownership is needed.

Alternatively, download hardwire-0.5.0.tgz and SHA256SUMS from this release.
Verify with shasum -a 256 -c SHA256SUMS (macOS) or
sha256sum --check SHA256SUMS (Linux), then install it in a separate project:

npm install /absolute/path/to/hardwire-0.5.0.tgz --ignore-scripts
node node_modules/hardwire/scripts/client-config.mjs

The tarball is distributed through GitHub; no npm package or MCP Registry listing
is published. Do not use npm install hardwire to obtain this project.

Checks exercise an official MCP client, all handlers, rejected inputs, a tampered
trace, and the command from a clean package install. Existing unit/CLI and real
HTTP integration suites also run in CI on Node 22 and 24. The release workflow
downloads the uploaded artifact, verifies its checksum and repeats the installed
MCP check before the draft is published.

This is experimental inspection tooling. It does not browse sites, authenticate,
execute submitted ports or certify inferred effects. Trace digests are consistency
checks, not signatures or proof of execution. Existing 0.4.x panel/trace formats
are unchanged.