Pixy Connect

Your coding agent installs Pixy, connects itself, and learns the job.

Pixy's setup is written for an agent rather than for a person. You copy a brief off your project page, paste it into Claude Code, Cursor, or whatever works on your codebase, and it does the whole install and reports back.

No card. The brief carries a one-time invitation, good for ten minutes.

You paste one brief. Your agent puts the tag on the site, mints a key of its own, and writes the workflow into your project. Nothing else changes hands.

What the brief tells your agent to do

Press Copy instructions on your project's setup page and paste. The brief names your site, carries your tag, and holds one live invitation.

1

Ask where the tag belongs

Development only, a staging build, or every build including production. It is the one thing the brief cannot know, so the agent is told to ask and wait rather than guess or answer from its own notes. Then it adds one script tag to the head of the build you named. No package, no build step, no config.

2

Sign up with the invitation

It trades the invitation for a key of its own and introduces itself by its model name and the machine it runs on, which is the name your panel shows from then on. It saves that key wherever your project already keeps credentials. The invitation is spent at that moment and never works again.

3

Install the workflow

It fetches the working contract with its new key, writes it where its harness finds skills, and adds one line to the instructions it loads every session: Pixy is connected, and here is where the key lives. Every session after that runs from that file and fetches nothing.

Approvals stay yours throughout. If the agent's harness gates a call it is told to let it ask, and to stop and say what was blocked if it is refused, rather than reaching for the settings itself.

Three keys, and only one of them opens anything

They are minted at different moments, for different holders, and no two of them can do the same thing.

The project key

The one in your script tag. It names the project and grants nothing, which is why it is safe sitting in public HTML.

The invitation

One use, ten minutes. The only thing you ever paste, and it dies the moment the agent trades it in.

The agent key

The agent's own credential for one project. It opens that project's edits and conversation, and nothing beyond them.

Your own right to edit is none of these. It lives in your browser, granted by clicking the black Connect Pixy button on your own site, and it is the one thing the agent cannot do for you.

What “Check Pixy” runs

Three words, and the agent runs a contract it already has on disk. It reads no documentation, and you copy nothing across.

Pull

It asks for everything open with its key: the tasks you packaged and the loose edits you never did. Pulling locks a task to that agent, and it works them oldest first, reading each thread before it builds.

Build

It implements each intent in source at the scope the record names. Real classes and real components, never an inline patch on the one instance you happened to click.

Ack

Every record is reported so it stops arriving: built, declined with a reason, or postponed. Loose edits get named into a task at the ack, so your history reads as packaged work either way.

Report

It closes the batch with a summary on the task's thread, saying what changed and where. That lands in your panel, and you answer there.

Reading is not resolving. An unacked batch simply arrives again on the next pull, so a session that dies mid-task loses nothing.

What a connected agent cannot do

Edit your site

It reads records and writes code. There is no call in the API that changes your running page.

Approve a browser

The right to edit is granted by you, on your own site, in your own browser. No key mints it and no agent can ask for it.

Reach past the project

An agent key opens one project's edits and conversation. It never touches your account, your plan, or another project.

Widen its own permissions

An agent editing its harness's settings to clear its own path reads as self-granting, so the brief forbids it and tells it to come back to you instead.

The instructions cannot go stale

Your agent works from its own copy of the contract and never re-reads the docs. Pixy remembers which version that copy is, and when the contract moves on, every response it gets starts saying so.

Then, and only then, it fetches the new one and writes it over the old.

Design on the real thing.

Get started

No card required ยท works on localhost, staging and production