Lightcms

What jonradoff/lightcms MCP server does

The jonradoff/lightcms MCP server connects MCP-compatible agents to LightCMS, a Go-based content management system designed for agent-controlled website operations. The repository describes an MCP interface with 106 tools and three prompt resources for managing site content and configuration.

Agents can work with reusable templates, custom fields, pages, folders, tags, snippets, collections, themes, assets, redirects, and multiple site forks. Content has version history, diff comparison, soft deletion, and revert support. Publishing can happen immediately or at a scheduled time, while approval workflows and contributor permissions can keep changes out of production until review.

The system also includes semantic search, generated llms.txt files, schema.org WebSite JSON-LD, analytics, maintenance reports, and a public read-only MCP endpoint for site visitors' agents. These capabilities make it relevant for teams that want both agent-operated content management and machine-readable site discovery.

How it works

The jonradoff/lightcms MCP server can communicate over local stdio or HTTP streamable transport. Local clients can connect directly to the running application, while remote agents can use OAuth 2.1 with PKCE. The README specifically names Claude Code, Claude Desktop, and other MCP-compatible clients as supported connection targets.

Content operations are exposed as MCP tools rather than requiring an agent to navigate the admin UI. The available operations cover individual edits as well as larger workflows, including Markdown, ZIP, CSV, and RSS/Atom imports. Import jobs provide status and cancellation operations, and repeated imports match content by full_path so an existing item can be updated instead of duplicated.

For safer agent workflows, LightCMS provides isolated forks with copy-on-write behavior, scoped API keys, sandbox-only keys, a session change ledger, rollback, and per-version provenance. Humans can review field-level differences before merging sandbox changes into live content.

Setup and configuration

The jonradoff/lightcms MCP server is part of the LightCMS application and is written in Go. The supplied material does not provide a verified package-manager command, binary invocation, Docker image, or complete configuration file, so installation details should be taken from the repository's current documentation.

Transport choice depends on the client and deployment model. Local use is described through stdio; remote use is available through HTTP streamable transport. Remote desktop-style connections can use OAuth 2.1 and PKCE rather than embedding a password. LightCMS also supports LIGHTCMS_EMBEDDINGS_PROVIDER=ollama for self-hosted semantic search, although the excerpt does not identify all configuration variables or state that this setting is required.

Tools and capabilities

The MCP surface described in the repository includes capabilities for:

• Creating, editing, publishing, unpublishing, deleting, restoring, and reverting content.
• Reading version history, comparing changes, and working with content forks.
• Managing templates, fields, assets, snippets, themes, collections, folders, tags, and redirects.
• Importing Markdown, ZIP archives, CSV files, and RSS/Atom feeds.
• Listing, starting, inspecting, and canceling import jobs.
• Managing comments, approval requests, and approval workflows.
• Reading analytics and maintenance information through the copilot and agent workflows.
• Searching site content semantically and exposing site structure, recent content, and theme configuration as MCP resources.

Limitations and notes

The excerpt contains inconsistent tool counts: the current description says 72 MCP tools, while the README describes 106 tools. The README's versioned feature notes also identify different subsets added in earlier releases, so consumers should verify the deployed version before relying on a particular tool.

The material does not include a license, a complete environment-variable reference, or a confirmed standalone installation command. It also does not establish whether external AI features, such as the Claude-powered chat widget, require a separate paid credential. Local Ollama embeddings are explicitly described as avoiding API credits, but that option does not document the requirements of every other AI-related feature.

The public MCP endpoint is read-only, while the management interface can write and publish content. Teams should therefore distinguish visitor-facing agent access from authenticated administrative operations when designing permissions and deployment boundaries.