We built MDCMS, a Markdown-first CMS for teams using AI agents

A developer has built MDCMS, an open-source, Markdown-first CMS that allows both human editors and AI agents to manage content through raw files or a database-backed interface. The tool includes a CLI that scans existing repositories for Markdown and MDX files, automatically inferring content types and syncing them into the CMS as drafts. MDCMS aims to solve the adoption problem for teams with existing content in repositories, enabling AI agents to safely work with content at scale while maintaining validation, permissions, and rollback capabilities.

Working with Markdown files for content is super convenient. As a developer, it is hard not to like it. Content is just files. You can search it, diff it, review it in PRs, move it around, and now also let LLMs bulk-manage it. Need to add metadata to 200 posts or migrate content from one structure to another? Files are very easy to reason about. But this does not scale that well for editors. Most marketing and content teams do not want to open a repo, understand folder structure, run commands, and wait for deployments just to change a page. They need drafts, preview, permissions, version history, publishing flows, rollback, localization, and all the other CMS things that are still very much needed. So the idea behind MDCMS was pretty simple: why not both? MDCMS is an open-source CMS that is database-backed and Markdown-first. The database stays the source of truth, so you still get the workflow and safety of a real CMS. But Markdown/MDX stays the working format, so developers and AI agents can still work with raw files when that is the better tool for the job. The part that matters a lot for brownfield projects is adoption. A lot of teams already have content sitting in their repository: blog posts, docs, landing pages, changelogs, MDX pages, marketing content. The problem is not starting from zero. The problem is turning what already exists into something editors can safely manage. With MDCMS, the goal is that you can run: mdcms init And the CLI walks through the existing repository. It scans for .md and .mdx files, groups content directories, detects locale patterns, parses frontmatter, infers content types, generates mdcms.config.ts , syncs the schema, and pushes the discovered content into the CMS as drafts. That is the adoption path I care about most: After that, the workflow looks roughly like this: For example: mdcms pull update content, metadata, links, translations, SEO fields mdcms push --validate The interesting part is not "AI writes blog posts." The interesting part is giving AI agents a safe way to work with content at scale, while still keeping validation, permissions, history, and rollback. For us, AI-native CMS means AI can work across multiple layers: | Layer | What AI can help with | |---|---| | Content | Markdown/MDX, metadata, links, translations, SEO fields | | Configuration | schemas, environments, locales, project settings | | Codebase | adapters, components, validations, missing workflows | | CMS actions | eventually controlled actions like users, roles, and permissions | That is the part raw files alone do not solve. And it is also the part many CMS dashboards are not great at, because the content is often trapped behind UI flows and proprietary APIs. MDCMS is our attempt to keep the useful parts of both models. Files when files are better. CMS when teams need a CMS. It is open-source and still early, so contributions are very welcome. If this problem sounds familiar, check the repo, try it on a small Markdown/MDX project, open an issue, or pick something from the roadmap. Website: https://www.mdcms.ai/ https://www.mdcms.ai/