Skip to content

MCP Server v2 — Code Mode, Skills, and Agent Governance #36078

Open
13 of 13 issues completed
Open

Description

@fmontes

Description

MCP Server v2 expands dotCMS's Model Context Protocol server so external AI agents (Claude, Cursor, ChatGPT desktop, etc.) can operate the full hybrid-CMS surface — content, pages, templates, workflows, and publishing — through a single curated entry point, governed by dotCMS's existing permission, audit, and logging systems.

Problem. v1's 5 hand-written tools cover content CRUD and search only. Operational asks now span pages, templates, workflows, navigation, and publishing. Hand-written tools don't scale to dotCMS's 600+ endpoints, and there is no skills/plugin distribution.

Solution.

  • Code execution pattern — two tools (search and execute) over an endpoint catalog represented as a filesystem (one file per endpoint, with schema + examples), avoiding the per-endpoint authoring tax and context bloat of one-tool-per-endpoint.
  • Curated endpoint whitelist — a dotCMS-maintained list of agent-safe endpoints with per-endpoint risk categorization (read / mutate / destructive), an internal intake process for additions, and versioning so each release exposes a known surface. Inherits dotCMS's existing permissions, audit, and logging.
  • Auth + tenant scoping — API-key flow with site/environment selection, scoping enforced at the MCP server on top of dotCMS's permission model.
  • Skills pack (Claude plugin) — proposed skills (dotcms-guide, dotcms-content-modeling, dotcms-page-building, dotcms-workflow, dotcms-search), distributed via npm install; marketplace submission best-effort.
  • Documentation & enablement — setup, whitelist/governance, and per-skill docs; operator enablement (first-time agent setup against dotCMS); self-service troubleshooting; v1 migration guide.

Critical-path spike. AI eval harness for v2 task performance — validate that agents reliably perform multi-step operational tasks (notably page creation) through search + execute. Timebox: 2 days; establishes the success-metric baseline.

Out of scope: in-product prompt-driven builder UI; dotCMS-branded chat/copilot/agent-builder in the admin; OAuth-based MCP authorization (and hosted MCP); new audit/observability surfaces; new dotCMS API endpoints; multi-LLM provider abstraction; replacing dotAI runtime features; a customer-facing agentic-app SDK.

Desired Outcome

  • An agent connected to v2 can compose a full workflow end-to-end (model content type → build page → transition workflow → publish) against a real dotCMS instance, under the curated whitelist and governed by dotCMS's existing permission model.
  • A dotCMS user installs the Skills pack in their MCP client and their assistant acts idiomatically against dotCMS with no further setup.

Target Personas

  • Developer teams
  • Content teams
  • DevOps teams
  • System administrators (dotCMS)

Activity

  1. added theissue type on Jun 10, 2026
  2. hegu-1 commented on Jul 13, 2026

    @hegu-1

    For a CMS MCP server with publishing/workflow reach, the governance surface is as important as the tool surface.

    The two-tool search / execute pattern scales nicely, but I would make sure the executable endpoint metadata carries enough policy information for agents and operators:

    • required permission/capability
    • side-effect class: read, draft, publish, delete, workflow transition
    • whether human approval is required
    • audit event id / correlation id
    • redaction or output-firewall behavior
    • deterministic workflow vs free-form agent action

    Otherwise a model sees 600+ endpoints as just "tools", while the organization sees very different risk levels.

    I have been thinking about this as an enterprise AI OS problem: models are swappable, but workflow, permission, memory, and audit planes need to be stable: https://github.com/hegu-1/enterprise-ai-os-architecture

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions