MCP server
Let an AI assistant build, validate, export and scaffold a Kirigami project through the Model Context Protocol.
Overview
@kirigami/mcp puts the Kirigami engine behind a Model Context Protocol server. An AI assistant that speaks MCP (Claude Code, Claude Desktop and others) can then build, export, validate and run scripts in your project the way the terminal does, and read structured results instead of parsing console output.
The server talks over stdio and works on the directory it is started in, exactly like kiri.
Setup
With the CLI installed in the project, no separate install is needed: point your MCP client at kiri mcp, started from the project folder.
npx kiri mcp
Or run the package directly from the client's configuration:
{
"mcpServers": {
"kirigami": {
"command": "npx",
"args": ["-y", "@kirigami/mcp"],
"cwd": "/path/to/your/kirigami/project"
}
}
}
The working-directory key depends on the client; what matters is that the server starts in the folder that holds kirigami.yaml. It also starts in an empty folder, where the assistant can scaffold a new project first.
Tools
| Tool | What it does |
|---|---|
kirigami_about |
A compact manifest of the project (package metadata, docs, tasks and scripts), to orient the assistant. |
kirigami_config |
The resolved kirigami.yaml and the active plugins. |
kirigami_validate |
Validate kirigami.yaml against the schema. |
kirigami_build |
Same as kiri build. |
kirigami_export |
Same as kiri export, with an optional output path. |
kirigami_list_tasks / kirigami_run_task |
List the build tasks, run exactly one. |
kirigami_list_scripts / kirigami_run |
List scripts/*.php, run one with arguments. |
kirigami_list_templates / kirigami_create_project |
List the official templates, scaffold a project from one. Never overwrites files. |
kirigami_site_blueprint |
Suggest a site tree, configuration and an example page. |
kirigami_search_docs / kirigami_doc_hints |
Search the project's documentation, or suggest the pages for a topic. |
Results are JSON: an operation reports success: false in its payload rather than as a protocol error, so the assistant sees the diagnostics.
There is no serve or watch tool: a long-running server doesn't fit a request/response protocol. Run kiri serve or the VS Code extension for a live preview.
Trust
kirigami_run and kirigami_run_task execute whatever the project's scripts and tasks do. Point the server only at projects you trust.