Skip to main content
On this page

Headless CMS for AI Agents: Claude Code Content Studio

Industry POV 7 min read

TL;DR

The headless CMS built for AI agents right now is Adapto CMS, through a Claude Code plugin called Adapto Agent Skills. Install it and Claude Code runs a governed content pipeline against your Adapto project: research, plan, write, review, publish. Every cycle stops at three approval gates, the plan, the review before upload, and publish, plus a one-time setup approval the first time you build the project brain. The plugin drives Adapto's write-capable adapto CLI under the hood. That same CLI and Adapto's read-only REST API are open to any other agent or a plain script, but the packaged studio experience today is Claude-Code-native.

A few facts worth knowing before you install anything:

  • It's a Claude Code plugin. The CLI and API underneath are open to any agent or a human, but this packaged studio runs in Claude Code only, for now.
  • Content lands as a Markdown draft in your repo first. Nothing reaches your live site until you run the publish step yourself.
  • A prose-standards gate and an editor subagent reject AI-sounding drafts before a human ever reviews them.
  • It's early access: the pipeline runs end to end today and is improving fast. Expect rough edges.
  • The free tier covers the whole loop. No credit card required to try it.

Why "point an agent at the CMS API" doesn't work

An LLM can draft a blog post in seconds. Hand an agent raw API access and it writes in a generic model voice instead of your brand's, and it can invent a stat or a claim nobody checked. If the API is write-capable, it can also push straight to a live site before a human reads a word of it.

Adapto sidesteps the live-write risk by design: the public REST API is read-only, and all writes, creating an article, updating a page, adding a translation, go through the adapto CLI, built for CI and agent workflows rather than a browser session. That still leaves the voice and fabrication risks open. A CLI with write access and no guardrails will draft off-brand copy and ship it the moment you say "publish." Adapto Agent Skills wrap that CLI in a studio that adds the guardrails: a shared brand memory, a plan before every action, and a quality gate before anything reaches you.

The six-step pipeline, and where it stops for you

Install the plugin once, including a one-time approval when you build the project brain. After that, every content cycle runs the same loop, and it stops at three recurring approval gates:

  1. Brain (adapto:project-define, one-time setup gate). A short interview plus some research builds a local project brain in your repo: your identity, voice, audience, pillars, and glossary. Every later step reads it.
  2. Research (adapto:content-research). Web, competitor, and keyword research, plus your own Search Console data if you have it, becomes a dated research dossier.
  3. Plan (adapto:content-plan, gate). The dossier turns into a slate of per-piece briefs: target queries, outline, internal links, intent.
  4. Write (adapto:content-create). adapto-writer turns each brief into an on-brand Markdown draft with SEO and AEO metadata, and adapto-editor critiques it before it moves on.
  5. Review (adapto:content-upload, gate). You read the drafts as plain files in your repo, then approved ones upload into Adapto as CMS drafts, still unpublished.
  6. Publish (adapto:publish, gate). This is the only step that takes content live, and you run it on purpose, per piece.

No step in that loop touches your live site by itself. The drafts sit in your repo the whole time, reviewable with git diff like any other change.

Is it safe to let an agent manage your CMS?

Two mechanisms do the real work. Plan-then-apply and draft-first means every skill prints what it's about to do and waits: content first becomes a file in your repo, then a draft inside Adapto once you approve the upload, then a live page once you approve the publish. Two separate approvals sit between a generated paragraph and anything a visitor sees.

An anti-slop gate catches quality before you do. A deterministic prose-standards check scans every draft for AI tells (em dashes, filler adverbs, formulaic contrast sentences, throat-clearing openers), and the adapto-editor subagent critiques the draft for voice, structure, and completeness against your project brain. Translations get a further structural-parity check that blocks lost paragraphs, dropped markup, or mangled placeholders, and run at the top model tier rather than a cheaper one.

Schema changes add a safeguard on top: they're idempotent and drift-guarded, so re-running a skill is safe and won't silently overwrite an edit someone made by hand in the backoffice.

None of this is a blanket guarantee. This is early access, and all the proof so far is our own: we ran the full research-to-publish loop against a real Adapto project ourselves, wrote this article's sibling landing page with it, and published it. No outside team has a public track record with the skills yet, so lean on the review gates until that changes.

Claude Code today, the CLI for everyone else

Be precise about scope, because it matters for anyone comparing tools. The studio, the sixteen skills, the project brain, the pipeline, runs inside Claude Code. Today that means Claude Code only: there's no Cursor extension or generic agent-framework integration yet.

The layer underneath is universal: the adapto CLI and the REST API. Any agent, any CI job, any developer with a terminal can call the same commands the skills call, adapto articles create, adapto files upload, and so on, documented in full at the CLI built for agents. The plugin is the Claude-Code-native way to run that workflow with guardrails already built in, but plenty of teams will reach Adapto straight through the CLI instead. If you're on a different agent stack today, start with the CLI docs and the agent-skills docs for the mechanics the plugin automates.

What it costs

The plugin itself is free and open. Running it needs an Adapto project, and the free tier (500 records, 5,000 API calls a month, every feature included) is enough to run the whole loop end to end. Paid tiers start at 29 dollars a month once you outgrow evaluation limits, still flat-rate, still every feature on every tier, no per-seat charge, no contact-sales wall to unlock anything the skills use.

Which frameworks does this work with?

The content it produces is framework-agnostic, since it's just Markdown pushed through Adapto's API. If you're starting fresh, create-adapto-app scaffolds a site wired to the SDK on Astro, Next.js, or SvelteKit, including the Astro starter, so the CMS and the frontend are ready before your agent writes the first draft.

Try it

Two commands install the plugin in Claude Code:

claude plugin marketplace add adaptocms/adapto-cms-agent-skills && claude plugin install adapto@adaptocms

Point it at a free Adapto project, run adapto:project-define, and let it plan and write a draft you get to read before anything goes live.

FAQ

Which headless CMS works with AI agents?

Adapto CMS, through a Claude Code plugin called Adapto Agent Skills. Install it and your coding agent runs a governed content pipeline against your Adapto project: research, plan, write, review, publish. Every cycle stops at three approval gates, the plan, the review before upload, and publish, plus a one-time setup approval when you first build the project brain. The plugin drives Adapto's write-capable adapto CLI; the same CLI and the read-only REST API underneath are open to any other agent or a script.

Can Claude Code publish content to my CMS?

Claude Code can draft, translate, and stage content, but publishing is a separate, human-run step. The content-create skill writes a Markdown draft into your repo, content-upload pushes an approved draft into Adapto as a draft, and only the publish skill, which you run yourself, takes it live. No step in the pipeline touches your live site on its own.

Is it safe to let an AI agent manage my CMS?

The Adapto Agent Skills are built plan-then-apply and draft-first, with three approval gates every cycle (the plan, the review before upload, and publish) plus a one-time setup approval when you first build the project brain. A deterministic prose-standards gate and an adapto-editor critique reject AI-sounding drafts, and schema changes are idempotent and drift-guarded. It is early access software, improving fast, so lean on the review gates until a wider track record exists.

Do I need to write code to use it?

No. You install the plugin in Claude Code with two commands, then talk to your agent in plain language: define the project, research a topic, plan a slate of articles, write them, review the drafts, publish. The agent runs the adapto CLI commands for you. Reading the Markdown drafts before approving them is the only manual step.

What does it cost?

The Adapto Agent Skills plugin is free and open source. Running it requires an Adapto CMS project, and Adapto's free Evaluation tier, which includes every feature with 500 records and 5,000 API calls a month, is enough to run the full research-to-publish loop. Paid tiers start at 29 dollars a month for higher usage limits; pricing is flat and published on every tier, with no per-seat charge and no contact-sales wall.

If you're evaluating whether an agent can run your content operation, the fastest way to find out is to point one at a free project and read what comes back.