@kitn.ai/mcp

The Model Context Protocol server for @kitn.ai/ui. It gives an AI coding harness the real component API, a scaffolder, a theme tool and a debug tool, so it builds with this library instead of guessing at it.

It runs locally, over stdio, with no state and no network calls of its own: your harness launches it as a child process when a session starts and stops it when the session ends. You configure it once; you never start it by hand.

Configure

The server is one program, so every harness config is the same two facts: launch npx -y @kitn.ai/mcp, over stdio, with no arguments. These are the mainstream harnesses; the full list of thirteen we have tested, each with its own file and key name, is on ui.kitn.ai/guides/for-ai-agents.

Claude Code — the CLI writes the config, or put it in .mcp.json:

claude mcp add kai -- npx -y @kitn.ai/mcp

{ "mcpServers": { "kai": { "command": "npx", "args": ["-y", "@kitn.ai/mcp"] } } }

Codex — ~/.codex/config.toml (global) or .codex/config.toml (project):

[mcp_servers.kai]
command = "npx"
args = ["-y", "@kitn.ai/mcp"]

VS Code — .vscode/mcp.json, or MCP: Open User Configuration for the user-level file. The top-level key is servers, not mcpServers, which is the usual copy-paste mistake:

{ "servers": { "kai": { "type": "stdio", "command": "npx", "args": ["-y", "@kitn.ai/mcp"] } } }

GitHub Copilot CLI — a different file from VS Code's Copilot Chat:

copilot mcp add kai -- npx -y @kitn.ai/mcp

Cursor, Windsurf, Cline, Zed, Gemini CLI — the same mcpServers shape Claude Code uses, in each tool's own config file.

Hermes — in config.yaml, or through its CLI:

mcp_servers:
  kai:
    command: "npx"
    args: ["-y", "@kitn.ai/mcp"]

hermes mcp add kai --command npx --args -y @kitn.ai/mcp
hermes mcp test kai        # exits 0 on a completed connect

Pi has no MCP in its core, by design — its README says to build CLI tools with READMEs, or add MCP through an extension. So on Pi, use the command line as a tool (npx -y @kitn.ai/cli doctor) and paste llms.txt for the API, or add MCP with an extension and launch npx -y @kitn.ai/mcp through it.

Tools

| tool | what it answers | |---|---| | component_reference | the real kai-* API: props, events, slots, the property-vs-attribute rule | | scaffold | generates a working chat surface wired to your backend | | theme | brands the kit from a color or a description | | debug | catches the classic wiring mistakes |

What it reads, and why that matters

It runs in your project's directory and resolves the @kitn.ai/ui you have installed, so the answers describe the API that project actually has, and it reports that version on initialize. The range it declares is checked against the kit in the workspace by the repo's verify:workspace-ranges guard, because a stale bound would point an agent at an API the app does not have.

Pinning

npx -y @kitn.ai/mcp fetches the latest each session. If you would rather the server match the kit your project pins, install it and point the config at the local binary:

npm i -D @kitn.ai/mcp

Not the command line

The kai command line — create, add, doctor, dev, compile, eject, validate, and a mcp verb that forwards to this server — is @kitn.ai/cli. It is a separate install on purpose: this package is the only one carrying the MCP SDK, so a developer who only wants kai add does not download it.