MCP Server

Krios exposes a Model Context Protocol server at /api/mcp/{projectSlug}. Agents (Claude, Cursor, custom agents) connect, list tools, and invoke them with a management or preview key.

Endpoint

GET /api/mcp/{projectSlug}             # streamable HTTP transport (SSE)
POST /api/mcp/{projectSlug}            # for tool invocations

Auth

Bearer with a management or preview key:

Authorization: Bearer <api-key>

The MCP route runs its own inline bearer check (SHA-256 hash + lookup, plus active / expiry / key-type / project checks). A delivery key is rejected with 403 wrong_key_type; sessions are not accepted. MCP enforces tenant/project scoping and key-type gating (management or preview). Each tool call also runs the same authorization gate as the REST API: the key's permissions[] must include the tool's action (read/create/update/publish/unpublish/delete; an empty list = all actions) — otherwise the tool returns api_key_permission_denied — and the call counts against the key's rate limit (returns rate_limit_exceeded when over). Per-site key scoping is not yet narrowed on reads.

Tools

Built-in tools the server registers:

Tool	Action
list_content_types	Return every content type with apiName + name + flags
get_content_type	Read one type with its full field list
get_entry	Fetch an entry by id
get_entry_values	Resolved field values for an entry + locale (rich text as raw + html + text)
search_content	Paginated list by content type + site + locale + status
create_entry	Create with contentTypeApiName, siteId, locale, treeParentId, fields
update_entry	Update with version for optimistic concurrency
publish_entry / unpublish_entry	Publish or unpublish a locale
delete_entry	Soft delete
get_tree	Flat node list for one site
list_sites / get_site	Site metadata

Each tool's input schema is published via the standard MCP tools/list call.

Resources

The server exposes read-only resources for context-loading workflows:

Resource URI	Returns
krios://schema/content-types	Full content-type schema as JSON
krios://sites	Site list
krios://tree	Combined global + per-site tree for the default environment

Entry data is not exposed as a resource — use the get_entry / get_entry_values tools instead.

Permissions + audit

Every tool call is gated by the key's permissions[] (per the action it performs) and counted against the key's rate limit, in addition to tenant/project scoping and key-type gating. A key whose permissions omit the required action gets an api_key_permission_denied reply; over-limit calls get rate_limit_exceeded. (An empty permissions array allows all actions. Per-site key scoping is not yet applied to reads.) Audit log entries record the call with the MCP key in metadata.mcpApiKeyId (and metadata.mcpApiKeyType). Use the Agent activity report to see how much content an agent authored.

Example client config

For Claude Desktop:

{
  "mcpServers": {
    "krios": {
      "type": "http",
      "url": "https://cms.example.com/api/mcp/demo",
      "headers": {
        "Authorization": "Bearer krios_mk_..."
      }
    }
  }
}

For Cursor / other agents — the same URL + bearer header with whatever transport they support.

Recommended scopes

• Read-only agents (RAG, summarization) — preview key with no write permissions.
• Drafting agents (compose entries for editor review) — management key with create + update on contentType, no publish.
• Auto-publish agents — management key with publish. Pair with the governance analyzer's skipped_workflow detector if your project has workflows.