Core concepts

A few nouns carry the whole product. Learn these and the rest of the documentation reads quickly.

Artifact

The unit of work. An artifact is a live app, a readable document, or a stored file, backed by a real git repository, with one address and one history.

An artifact has a type:

app renders and runs as a web page, built from HTML or React.
markdown renders as a readable document you can edit in place, in a rich editor.
Anything else is an asset: stored, versioned, and readable by agents, without a preview. The hub filters by all three.

Its address looks like app.agentgrid.io/artifacts/Ez-CDDsmYLaApCIqpPv5bQ — the workspace, then the artifact's own id. That id is how every tool refers to it. The easiest way to get the real address is Share, which shows the link with a copy button.

Team

Artifacts belong to a team, not to a person. You may belong to several, and you switch between them from the team name in the top left, which also shows your role and the team's plan. Whatever you create lands in the team you are currently in.

People hold a role in a team: Viewer, Editor, Admin, or Owner. You see and change roles under Share → Manage access. See Team roles.

The team is the boundary of access. Everyone in a team can open every artifact in it — a role changes what they can do, not what they can see. Anything you want kept private belongs in a team of its own, and anything meant for a different audience belongs in a team made for them, with the artifact duplicated across.

Agent

Any AI connected to AgentGrid.io with its own identity. An agent is not a shared login. It has a name, it acts on behalf of one human, it holds only the capabilities that human granted, and it does not use a team seat.

How they get in depends on what they can do:

Chat assistants. The ordinary chat in ChatGPT or Claude, and anything running in a browser. They connect through the official connector in Claude, or the official plug-in in ChatGPT, installed from that assistant's own directory.

Coding and autonomous agents. Claude Code, Codex, Cursor, OpenClaw, Hermes, Claude Cowork, worker packs, desktop agents. One command connects them: npx @animaapp/cli@latest login.

Anything in a browser is a chat assistant, whatever it is called. See How agents connect.

The four capabilities:

Capability	What it allows
`read`	Open and read artifacts in scope
`write`	Change artifact content
`share`	Give others access
`publish`	Deploy an artifact to a public URL

An agent can hold fewer than its human has. It can never hold more. Every request is checked against the human's live permissions, so a change takes effect immediately rather than at some future expiry. Agents are listed with their capabilities under Share → Manage access, where they can also be removed.

Claiming

An AI that reaches AgentGrid.io without connecting can still create an artifact, with no account on either side. Those artifacts are read-only, owned by nobody, and expire 24 hours after creation. Claiming moves one into your team, unlocks editing, and stops the expiry.

A connected AI does not produce one. It already has an identity in your team, so what it makes lands there directly. See What is claiming.

Git

Every artifact is a git repository, and that is what makes the rest work: real commits, real history, real attribution.

Agents change source files by committing. People edit markdown artifacts directly in the app, and those edits land in the same history as "Manual Edit" entries. Any earlier version can be restored from the History panel.

How they fit together

An app artifact's History panel, with human and agent changes side by side

Two words that are not the same

Sharing is giving someone access to the artifact: the team link, a team invite, or an agent invite, all from the Share button. Publishing deploys the artifact to a public URL that anyone on the internet can open, with no account. They are separate, deliberate actions, and an agent should never publish unless it was asked.