はじめにGet started

Publish with an agent

Ask your AI agent for a prototype, a report, a dashboard or a small site, and Kazari gives it an address on your organisation’s domain. You do not need a repository, a build step or a hosting account. This guide is for anyone who works with an agent, whether you write code or not.

You need a Kazari instance. Your administrator will give you its address, which looks like art.example.com in the examples below. If you are the administrator, start with Install.

Connect your agent#

Kazari is an MCP server at https://mcp.art.example.com/ (your instance’s mcp. address). Any agent that can use a remote MCP server over Streamable HTTP can connect, whether it runs in the cloud or on your own machine. It either signs you in through OAuth or sends an API key as an Authorization: Bearer kzk_... header.

Cloud agents#

Agents that run in a web or desktop app add Kazari as a custom connector (some apps call it a remote MCP server):

1. In your agent’s settings, add a custom connector with the address https://mcp.art.example.com/ and connect.
2. The app sends you to sign in. If your organisation has set up single sign-on, use your work account; otherwise Kazari asks for an API key, which your administrator can issue for you.
3. Kazari shows a consent page naming the app, your account and what it may do. Approve it once.

In claude.ai, for example, custom connectors are under Settings → Connectors, and on Team and Enterprise plans an owner may need to add the connector for the organisation before members can connect it. A cloud agent reaches Kazari from the internet, so your instance needs a public https address.

Self-hosted agents#

Agents that run on your own machine or servers, such as command-line coding agents, take the server in their configuration. Many take a JSON entry along these lines; check your agent’s documentation for where it goes:

{
  "mcpServers": {
    "kazari": {
      "url": "https://mcp.art.example.com/",
      "headers": { "Authorization": "Bearer kzk_..." }
    }
  }
}

Leave out headers to sign in through OAuth instead, if your agent supports it. Claude Code, for example, adds the server with one command and opens your browser to sign in the first time it uses it:

claude mcp add --transport http kazari https://mcp.art.example.com/

Your instance also serves a skill with its conventions, limits and recipes. Agents that load skills over MCP find it on their own; others can call the get_guide tool. To install it for all your projects in Claude Code, download it with your API key (or run kazari skill with the CLI):

mkdir -p ~/.claude/skills/kazari
curl -fsSL -H "Authorization: Bearer kzk_..." https://api.art.example.com/v1/skill \
  -o ~/.claude/skills/kazari/SKILL.md

Ask for a page#

Say what you want and that it should go on Kazari. Naming a slug is optional, but it decides the address and cannot be changed later.

The agent replies with the address, for example Published https://signup-flow.art.example.com/ (version 1). Apps that support MCP Apps also show a card with links to the page and the version.

Versions and addresses#

Every publish creates a new version, and old versions are never changed or thrown away:

• https://signup-flow.art.example.com/ shows the published version, usually the newest.
• https://signup-flow--v2.art.example.com/ shows version 2, always.

Share the main address with people who should see the latest work. Put a version address in feedback, sign-off or a ticket when it matters exactly what someone looked at.

Share safely#

• Public pages appear in the listings that other Kazari users see. Unlisted pages are left out of other people’s listings but are still open to anyone with the link.
• Unpublished takes a page offline at once (visitors get “not found”) without deleting anything. Ask your agent to make it public or unlisted again to bring it back.
• Search engines are told not to index pages unless you ask for the page to be indexable.

Tidy up#

Pages you no longer need can retire themselves. Ask your agent to mark a page stale with a grace period (“retire it in two weeks”). The page keeps working until the grace period ends, then:

1. it is soft-deleted: offline, but it can be restored for 30 days (by default; ask your administrator);
2. it is purged: the files are gone, and old links say “gone” for 90 days;
3. the slug is freed and can be used again.

Changed your mind during the grace period? Ask your agent to unmark it, or publish a new version, which clears the mark too. Deleting a page outright is possible too; the agent should check with you before a permanent delete, and apps that support it ask you to confirm.

What works in a page#

Kazari serves static files: HTML, CSS, JavaScript, images, fonts and data files. Your agent learns these rules from the tool descriptions, but they help when you ask for something ambitious:

• One index.html is enough. It is served at the root address. Extra files (styles, scripts, images, JSON data) are fine, and links between them should be relative (about.html, not /about.html) so they also work on version addresses.
• Libraries from the common CDNs work. By default scripts may load from cdnjs, jsDelivr and unpkg, styles from Google Fonts, and images from any https address. Other sources are blocked by the browser.
• Data travels with the page. A page can fetch its own files, such as data.json, but not other websites or APIs. Put the data in the page or next to it.
• No server-side code. There is nowhere to run PHP, an API route or a database, and forms cannot save anything on the server. For a prototype, fake the behaviour in JavaScript.
• Each page is its own site. localStorage and cookies are private to the page, so a prototype can remember state without seeing anyone else’s.
• Single-page apps that route in the browser need a setting so that every path serves index.html: ask your agent to set the metadata kazari.spa to true.
• Size. By default a version can hold up to 50 MiB and 2,000 files, and one file up to 25 MiB. Above about 2 MiB, the agent is told to send files in chunks, so large pages work too.

Your instance’s limits and allowed sources may differ; the tool descriptions your agent sees always state the real ones.

When something goes wrong#