If you have connected Cursor, Claude Code, or ChatGPT to a CMS in the last year, you have already voted in an informal standards election. The winner is Model Context Protocol (MCP): a way for agents to call typed tools instead of guessing from HTML exports or undocumented REST.
Sanity ships a hosted MCP server. Contentful’s remote MCP server went generally available in 2025–2026. Strapi enables MCP in config. WordPress published an official MCP Adapter that maps the Abilities API to tools. Storyblok and Hygraph document MCP against their Management APIs.
That convergence is real. It is also incomplete.
MCP standardizes the cable. It does not standardize what problem you are solving when you plug an agent into a CMS. Vendors still bundle three different jobs into one marketing slide:
- Ops agents — change structured content from an IDE or chat client.
- Editor assistants — suggest copy inside the admin UI.
- Public/agent layer — tell the wider world (and other people’s models) what your organization is allowed to claim.
Most comparisons stop at “we have MCP too.” For builders shipping sites, the interesting question is whether your stack makes those three surfaces explicit, governed, and composable—or whether you get a translation button and a prayer.
This article maps how leading CMS platforms implement AI workflows today, compares them directly, and explains what Birkly optimizes for. Features change quickly; a disclaimer at the end applies.
The four workflow shapes you will actually encounter
Before naming vendors, it helps to name shapes. Every “AI CMS” story is some mix of these:
| Shape | Who | Trigger | Output |
|---|---|---|---|
| External MCP agent | Developer in Cursor / VS Code | Natural language → tool calls | CRUD on entries, schema, assets |
| In-admin assistant | Editor in Studio / admin UI | Sidebar or inline “generate” | Suggestions → human accept → save |
| Server automation | Your code, webhooks, “functions” | Publish event, cron, pipeline | Drafts, patches, translations at scale |
| Public agent surface | Visitor or third-party model |
/.well-known/*, browser tools |
What the world may read or safely do on the site |
Sanity is strong on all four if you pay attention to credits and APIs. Contentful leans heavily on the second and third for marketers. WordPress is strongest on external agents if you assemble adapter + ability plugins—and weakest on structured content for agents to reason about. Sitecore and Kontent.ai ship packaged in-admin agents with workflow gates.
If your team only evaluates “MCP checkboxes,” you will optimize the wrong layer.
Sanity: schema-aware automation for people who write code
Sanity’s MCP server is hosted at mcp.sanity.io. Agents run GROQ, patch documents, manage releases, and explore schema without you pasting TypeScript definitions into the chat.
That is table stakes for developers. Sanity’s differentiated bet is Agent Actions—programmatic, schema-aware LLM operations (generate, transform, translate) you invoke from Functions, webhooks, or migration scripts. By default they create drafts; publishing requires an explicit forcePublishedWrite.
Editor-facing, Content Agent is conversational: find gaps, propose bulk edits, and confirm before apply (drafts or Content Releases). AI Assist covers quick field-level help inside Studio.
Typical Sanity workflow
- Developer connects Cursor → MCP → migrates schema or bulk-patches 400 legacy documents.
- Marketer publishes an article → Function fires → Agent Action translates → release branch → human review → publish.
Where Sanity wins: treating the content model as the contract for both humans and machines; releases and audit as first-class citizens for AI output.
Where it costs you: you are in Sanity’s ecosystem and pricing; local/self-host is not the positioning. Automation depth assumes engineering time to wire Functions + Actions.
Contentful: no-code AI Actions chained into marketing workflows
Contentful optimized for marketing operations at scale. AI Actions let teams define prompt + LLM pairs for translation, SEO rewrites, QA checks, and brand consistency—invoked on whole fields (not arbitrary text selections). Automations chain triggers to actions; Workflows add approval stages.
The remote MCP server exposes the same universe to external agents: entries, content types, releases, taxonomy, and even create_ai_action / invoke_ai_action. That is unusually complete for an enterprise SaaS.
Typical Contentful workflow
- Writer moves entry to “Ready for SEO.”
- Automation invokes an AI Action on meta title and description fields.
- Workflow routes to editor approval.
- Cursor, in parallel, runs
semantic_searchacross the space for a refactor project.
Where Contentful wins: marketers can own automations without deploying Node functions; governance maps cleanly to spaces, roles, and workflows.
Where it costs you: advanced AI is tiered (Enterprise add-ons matter); agents act as the connected user (proxy permissions), which is fine until you need dedicated agent identities with different blast radius; field-level bulk actions are powerful but awkward when you need entry-level reasoning.
WordPress: MCP as a plugin sandwich
WordPress did something principled: the MCP Adapter is infrastructure, not a content feature. It exposes Abilities as MCP tools. Actual capabilities arrive via plugins—Postnova, Fyrins, WDS MCP Content Manager, and others.
Typical WordPress workflow
- Install MCP Adapter + a content ability pack.
- Create a dedicated WordPress user with constrained capabilities + application password.
- Point Cursor at
wp-json/mcp/.... - Agent drafts posts; human reviews in Gutenberg—or does not, if you were careless with capabilities.
Where WordPress wins: enormous installed base; you can expose exactly the abilities you want; dry-run patterns exist in newer plugins.
Where it hurts agents: content is often blocks and HTML soup, not a strict schema. Agents spend tokens inferring structure. Governance is your problem: capability matrix, plugin choice, and whether the agent user can delete_post. There is no universal “drafts only unless forced” semantic across plugins.
Enterprise “agentic CMS” (Storyblok, Hygraph, Kontent.ai, Sitecore)
A second wave sells agents inside the product:
- Storyblok Agents — chat over stories, scheduled audits, user-confirmed writes; FlowMotion (n8n) for longer automations on Enterprise.
- Hygraph AI Agents — agents bound to workflow steps (translate at step 3, SEO at step 4) with side-by-side diff and per-field revert.
- Kontent.ai — “expert agents” (compliance, SEO, modeling) orchestrated with workflow traceability; some agents use dedicated permission models rather than pure user proxy (industry analysis).
- SitecoreAI — large libraries of prebuilt agents, brand context, no-code flows across brief → asset → publish.
Typical workflow (Hygraph-style)
Entry enters “Localization” step → Translation Agent runs → editor sees field-level diff → approves → next step.
Where they win: non-developers get repeatable automation; approvals are visual; procurement loves “agents included.”
Where they cost you: platform lock-in, implementation services, and opaque credit models. External IDE agents become a second workflow unless MCP is kept in parity with in-product agents.
Strapi and the self-hosted headless middle
Strapi advertises a built-in MCP server (token-scoped tools) and Strapi AI for generating content types from prompts or designs—developer acceleration more than marketer orchestration.
Where Strapi wins: you host it; MCP permissions map to admin API tokens; attractive if DevOps is already yours.
Where it trails Contentful/Sanity: review workflows and native automation layers often sit behind Enterprise tiers; you may assemble more pipeline yourself.
What developers should demand (regardless of vendor)
After reading a dozen vendor pages and MCP tool manifests, these requirements survive hype:
- Schema is the agent contract. If the CMS cannot tell an agent field types, relations, and validation rules, you will ship silent data corruption.
- Drafts by default for machine writes. Sanity’s Agent Actions embody this; WordPress often does not unless you enforce it.
- Separate identities for automation. Personal OAuth connectors vs service accounts with capped roles beat one shared API key on an admin user.
- Approvals matched to channel. In-admin queue for casual editors; client-side tool confirmation for IDE agents; workflow stages for enterprise—but document which one applies where.
-
Procedures, not only tools. Agents fail on
{for each entry in blog}vs detail-page URL context. Playbooks beat raw CRUD. - Public truth distinct from admin ops. If your site makes factual claims, models should read a governed, versioned surface—not scraped marketing copy.
Most stacks nail #1 partially and #2–#6 unevenly.
Birkly: three layers, one connector
Birkly is a self-hosted CMS—flat-file content, no runtime npm dependencies on the core, AI wired in from day one. Most vendors excel at one slice (hosted content API, marketer automations, or WordPress plugins). Birkly ties admin MCP, public agent context, and live-site tools into one stack for teams that ship sites themselves.
Admin MCP (Settings → AI)
Paste a connector URL into Cursor, Claude, or ChatGPT. OAuth signs in with your permissions, or a service agent with a fixed role you control.
From that one connection an agent can:
- Model and fill collections (structured fields, validation, media ids—not block soup).
- Edit
project/HTML, CSS, and JS beside the CMS (write_project_file,get_project_info). - Follow playbooks baked into the session (templating, forms, fieldtypes) so fewer “obvious” agent failures.
- Run Agent Guide and full admin actions (
cms_admin_action, plus Commerce/Automation tools when plugins are on).
Sanity and Contentful offer excellent hosted MCP to their content APIs—ideal when your frontend is a separate Next.js repo and marketing owns release trains in their cloud. Birkly’s MCP is built for the case they do not optimize: one external agent session that models content, writes templates, and publishes a working site. You need a public HTTPS /mcp on your instance; you keep the data and the model bill.
Built-in admin chat uses your API keys (including Ollama). MCP writes confirm in the AI client; optional Birkly approval queue covers in-admin chat.
Agent Guide (what AI may say about you)
Most CMS AI features optimize production (more copy, faster translation). We added Agent Guide—a published JSON index with:
- Sources of truth (narrative, proof, reference roles)
-
Claims with evidence URLs (only
supportedclaims appear in public/MCP views) - Products with typed outbound links (CMS entry, marketplace, reviews)
Public: /.well-known/agent-guide.json. MCP read tools: get_agent_guide_context, claim/product graphs. AI Assist in admin suggests registry entries and cross-links; nothing publishes until a human clicks Publish.
Publish /.well-known/agent-guide.json: sources of truth, supported claims with evidence URLs, products with links. MCP read tools expose the same snapshot; drafts never leak. Competitors can approximate this with metadata and enterprise compliance agents; none ship it as a first-class, publish-gated well-known file in the core product.
WebMCP (what a visitor’s AI may do)
Separate from admin MCP, WebMCP exposes visitor-scoped tools on the live site—contact forms, cart actions, RSVP—off by default, always with visitor confirmation. Admin MCP never touches the public browser surface.
Visitor-scoped tools on the public site (forms, cart, RSVP)—off by default, confirm in the browser. Admin MCP never substitutes for this layer.
One prompt, one site (“bakery” smoke test)
| Step | Contentful / Sanity default | Birkly |
|---|---|---|
| Content model | Studio / content model UI | MCP create_collection
|
| Seed content | Entries in UI or MCP | create_entry |
| Frontend | Separate app repository |
write_project_file under project/
|
| Contact form | Custom app or integration | Collection + {entry for} + playbooks |
| Agent instructions | Schema in MCP | Schema + playbooks + project sandbox |
If your frontend is already a separate framework repo, Sanity or Contentful are natural fits. If the agent should own content + templates + forms in one place, Birkly is the straightforward match.
Simple scorecard
Ratings are blunt: Strong / Good / Weak for the job described in the row—not a count of MCP tool names.
| Job | Sanity | Contentful | WordPress + MCP | Birkly |
|---|---|---|---|---|
| External agent ships a full site (model, templates, forms, publish) | Good | Good | Weak | Strong |
| External MCP to a hosted content API (releases, huge spaces, SaaS ops) | Strong | Strong | Good | Good |
| Marketer no-code AI workflows (translate/SEO chains in UI) | Good | Strong | Weak | Good |
| Structured content agents can trust | Strong | Strong | Weak | Strong |
| Self-host + bring your own model | Weak | Weak | Strong | Strong |
| Published machine-readable org truth (claims, evidence) | DIY | DIY | DIY | Strong |
Contentful and Sanity win when marketing runs AI inside a big SaaS content operation. WordPress wins when the site is already there. Birkly wins when you connect Cursor to your CMS and expect a working site, not only API patches in another repository.
Bottom line
MCP is the default wire. The product question is what the agent is allowed to touch.
Headless SaaS CMSs are excellent content databases with MCP. Birkly is a site system with MCP: collections, project/, playbooks, Agent Guide, optional WebMCP—self-hosted, any model you plug in.
If that is the job you are hiring an agent for, the comparison is simple. If you need a global marketing automation suite in a vendor cloud, buy the suite that was built for that job.
Further reading
- Birkly docs — AI overview
- MCP connections
- Agent Guide workflows
- Sanity MCP · Contentful MCP · WordPress MCP Adapter
Disclaimer — snapshot date 16 November 2026 (Europe/Copenhagen)
CMS AI features change monthly. This article reflects public documentation and product positioning as of 16 November 2026. Names, tiers, MCP tool lists, and GA dates may have shifted by the time you read this. Verify against each vendor’s docs before architectural commitments.
Birkly is developed by the team at github.com/Raaakete/Birkly.
Top comments (0)