MCP

Kide includes a local stdio MCP server for AI-assisted, schema-aware content editing. Agents can inspect collections, manage drafts and translations, and update asset metadata without using the admin UI.

Start the MCP server

pnpm cms:mcp

Most clients start it for you. Configure them to run this command from the Kide project root.

Schema changes are picked up automatically

The server watches cms.config.ts, the collections/ directory, and the generated API. When any of them change on disk, the next tool call reloads the project in a fresh worker process — no restart or reconnect needed. After editing the schema, run pnpm cms:generate && pnpm cms:push (or keep pnpm dev running, which does both) so the generated API and database match before using the new collection. If a reload fails — for example a type error in cms.config.ts — tool calls return the load error; fix the file and call again.

Codex

From the project root:

codex mcp add kide -- pnpm cms:mcp

For a project-scoped setup, create .codex/config.toml:

[mcp_servers.kide]
command = "pnpm"
args = ["cms:mcp"]
cwd = "/path/to/your-kide-project"

[mcp_servers.kide.env]
KIDE_MCP_USER_ROLE = "editor"
KIDE_MCP_USER_EMAIL = "editor@example.com"

The env table is optional — it runs the server as an editor-like actor instead of the defaults listed under Access rules actor.

Claude Code

From the project root:

claude mcp add --transport stdio kide -- pnpm cms:mcp

Open Claude Code and use /mcp to confirm that kide is connected.

Generic client config

{
  "mcpServers": {
    "kide": {
      "command": "pnpm",
      "args": ["cms:mcp"],
      "cwd": "/path/to/your-kide-project"
    }
  }
}

What it exposes

pnpm cms:mcp runs kide mcp, a stdio server registered as kide-cms. Tools:

Tool	Does
kide_list_collections	List all collections and their high-level metadata
kide_describe_collection	Fields, value shapes, and options for one collection
kide_list_documents	List documents with filters, search, sort, locale, availability, and status
kide_count_documents	Count documents with the same filters
kide_get_document	One document by id, optionally overlaying a locale and selecting status
kide_create_document	Create a document from declared schema fields (a draft in draft-enabled collections)
kide_update_document	Update declared schema fields on a document
kide_publish_document	Publish a draft-enabled document
kide_unpublish_document	Unpublish a draft-enabled document
kide_schedule_document	Schedule publish (and optionally unpublish) at ISO timestamps
kide_get_translations	All stored translations for a document
kide_upsert_translation	Create or update one locale’s translation (translatable fields only)
kide_list_assets	Asset records, optionally scoped to a folder
kide_update_asset	Safe asset metadata: alt text, filename, folder, focal point
kide_list_asset_folders	Asset folders

Resource: kide://model returns the machine-readable content model — the same manifest pnpm cms:describe writes to .kide/model.json (see Migrations).

There are no delete tools.

Safety defaults

Create and update tools accept only declared schema fields; system fields are ignored, and publishing is always explicit. Draft-enabled collections create drafts by default, so publish only after review.

Auth collections, such as users, are blocked for MCP mutations by default. Enable them only when you intentionally want agent access:

KIDE_MCP_ALLOW_AUTH_COLLECTIONS=true pnpm cms:mcp

Access rules actor

MCP calls include an actor, so collection access rules still receive context.user. Defaults:

Variable	Default
KIDE_MCP_USER_ID	mcp-local
KIDE_MCP_USER_ROLE	admin
KIDE_MCP_USER_EMAIL	mcp@local

To run as an editor-like actor:

claude mcp add --transport stdio \
  --env KIDE_MCP_USER_ROLE=editor \
  --env KIDE_MCP_USER_EMAIL=editor@example.com \
  kide -- pnpm cms:mcp

Content workflow

The server uses the same local database as your Kide app (data/cms.db on Node unless CMS_DATABASE_URL is set). For production edits, work against local or staging content, review drafts and previews, then apply only approved changes. Do not routinely replace a production database with an AI-edited copy.