This chapter surveys the wave of AI-forward CMS products and the AI layers bolted onto established platforms during 2025–2026. We separate the structured-content / headless camp (Sanity, Contentful, Storyblok, Payload, plus agentic newcomers Cosmic, Kontent.ai, Contentstack) from the visual web-builder camp (Webflow, Framer, Builder.io). For each we record what is genuinely shipped versus roadmap or marketing, then consolidate into a capability matrix. The recurring thesis: schema-aware, structured content is what makes AI useful in a CMS — the platforms that already had strong content models are the ones whose AI features actually work, and the new battleground is the agent, exposed over the Model Context Protocol (MCP), not the chatbot.
Content
How to read "AI-native" in 2026
By 2026 nearly every CMS claims to be "AI-native," so the label is nearly meaningless on its own. A more useful lens is where the AI sits and what it can touch:
Inline writing assist — generate/rewrite/translate text inside a single field. Table stakes since 2023; everyone has it.
Schema-aware actions — AI that reads your content model and produces valid structured documents (correct field types, references, validation), not just free text. This is the genuine 2024–25 advance.
Workflow automation — AI steps triggered by content events (publish, status change), often metered and templated.
Agentic operations — an autonomous loop that can query, mutate, migrate, and audit content across many documents, increasingly exposed to external agents (Claude Code, Cursor) via MCP.
Generative build — for visual builders, generating whole pages/sites/components from a prompt.
The structured-content vendors compete on (2)–(4); the visual builders compete on (5). The honest reading of the market is that (1) and (5) are largely shipped and impressive, (2)–(3) are shipped but uneven, and (4) is where the most marketing outruns the product.
The structured-content / headless camp
Sanity — the most aggressive structured-content AI play
Sanity has gone furthest in repositioning itself, rebranding around a following an , with its Spring Release 2025 described as its biggest ever (PRNewswire, 2025; CMS Critic).
"Content Operating System for the AI era"
$85M Series C
What is actually shipped:
AI Assist — the original inline feature: schema-aware field generation, translation, and image-prompt assistance configured with "Assist Instructions" embedded in the schema.
Agent Actions API — the more important primitive. Five action types: Generate (new schema-valid content), Transform (modify while preserving structure), Translate, Prompt (raw LLM call using Sanity content), and Patch (structured field updates). Crucially, outputs are validated against your Studio schema. A June 2025 change made Agent Actions create drafts rather than mutate published documents by default — a meaningful safety improvement, and a tell that early versions were unsafe (Sanity changelog, June 2025).
GROQ integration — actions can template live values from GROQ queries; GROQ filters can gate which content changes trigger automations.
Hosted MCP server — first-party, lets Claude Code / Cursor run GROQ queries, manage releases, and patch documents with schema awareness. Sanity reported analyzing 1.5M agent tool calls to this server between Sept 2025 and April 2026 across thousands of organizations — a rare usage datapoint rather than a press claim (Sanity, content-ops report 2026).
Functions & Blueprints — event-driven serverless compute to wire Agent Actions into workflows.
Assessment: Sanity's AI is the most credible in the structured camp precisely because it leans on its existing strengths (schema + GROQ). The "operating system" framing is marketing, but Agent Actions and the MCP server are real, documented, and metered.
(As of 2026-06-05.) Sanity kept iterating in May–early June 2026 (per its official changelog, Verified): it launched Agent Context Insights (v0.4.0 May 4 → opt-in telemetry v0.5.0 May 5 → simplified classification v0.6.0 May 13) — a new product line that captures and classifies agent conversations (success score/sentiment via a scheduled Sanity Function), i.e. observability of the agents, a genuinely new layer. Its MCP server added MCP-managed schemas and a deploy_studio tool (v2.19.0, May 8), a give_feedback tool (v2.20.0, May 29) and a June 2 schema-deploy fix; Scheduled Functions hit GA (May 7); and Studio init gained coding-agent "skills" (v5.29.0, June 2), mirroring the Contentful Skills pattern below.
Contentful — AI Actions + Studio
Contentful split its AI story into two products:
AI Actions — generative AI embedded in workflows with pre-built templates (translation, SEO optimization, alt-text, grammar). Uses Variables and External References to inject brand voice, locale, and third-party data (e.g., Shopify) for contextual awareness. It is sold as a paid add-on, metered in "AI Consumption Units" based on word output with pay-as-you-go overages (Contentful, 2025). This consumption-unit model is notable — it makes AI a line item rather than an included feature.
Studio — a visual workspace for assembling experiences with reusable templates, content binding, and design controls; also an add-on. Studio is closer to the visual-builder camp and reflects Contentful's push beyond pure headless.
Pricing context: Platform tiers are Free ($0), Basic ($300/mo), and Premium (custom); Personalization, AI Actions, and Studio are all separate add-ons (Contentful pricing, 2026). Assessment: AI Actions is solidly shipped and the template library is practical, but the everything-is-an-add-on packaging means real-world cost is hard to predict.
(As of 2026-06-05 — two material updates.)(1) Contentful Skills (released May 20, 2026): a free, open-source package that teaches AI coding agents (Cursor, Claude Code, GitHub Copilot, and ~a dozen others) how to build on Contentful, shipped as installable skill packs (contentful-guide routing, contentful-nextjs, contentful-migration, plus a personalization skill). This is a developer-acceleration/copilot play, not an autonomous agent. (2) Salesforce acquisition: on June 1, 2026 Salesforce announced a definitive agreement to acquire Contentful (pending, not closed; expected to close ~Q3 of Salesforce FY2027; terms undisclosed; The Next Web). The stated intent is to make Contentful Agentforce's native content-orchestration layer. Implication (inference): Contentful's roadmap independence and pricing are now uncertain — anyone selecting Contentful for new work should weigh acquisition-integration risk. (Skills + acquisition = Verified; roadmap impact = Inference.)
Storyblok — AI tools via "Labs"
Storyblok launched Storyblok Labs in January 2025 as an opt-in hub for early-access AI features, free to toggle on/off (Storyblok press; Enterprise Times, Jan 2025). Shipped items:
Ideation Room (beta) — a collaborative pre-publication workspace with built-in AI text commands: Complete, Shorten, Extend, Rephrase, Summarize, Simplify, Adjust Tone/Emojify, Translate. Available across all plans as public beta.
Concept Room (preview) — a visual whiteboard for designing project/content structure.
AI Translations — one-click whole-page or per-field translation.
Bring-your-own-model — editors can pick OpenAI, Google Gemini, Perplexity, or Anthropic Claude and choose the model per use case, which is a differentiator in editorial control.
Assessment: Storyblok's approach is the most editor-centric and the least "agentic." The Labs framing is honest — these are explicitly experimental. Strong on writing assist and provider choice; light on schema-aware mutation or autonomous agents.
Payload — the developer/RAG angle
Payload (open-source, Payload 3.0 is Next.js-native) positions AI as a framework concern rather than an editor feature:
RAG-ready by design — Payload markets itself as the CMS that puts retrieval-augmented generation "on easy mode," with control over chunking and auto-embedding of content into vectors out of the box (Payload enterprise docs, 2025). This is a genuinely different positioning: the CMS as the retrieval substrate for your own AI apps.
Official MCP plugin — exposes Payload collections to external agents.
Community AI plugin (ashbuilds/payload-ai) — the most-used third-party plugin, tested against Payload v3.38.x, adding a writing assistant, image generation, summarization, keyword suggestion, and field-level generation. Requires an OpenAI key; optionally supports Anthropic and ElevenLabs. Notably, the marquee generative features are largely community/plugin, not first-party core.
Assessment: Payload's honest story is "we give you the primitives (vectors, MCP, hooks) to build AI features," not "we ship a polished writing assistant." Best fit for teams building AI products on their content rather than buying an editorial copilot.
The agentic newcomers and incumbents repositioning
A cluster of vendors are explicitly chasing the "agentic CMS" label:
Kontent.ai — launched its Agentic CMS in October 2025, claiming "world's first," then added Expert Agents (announced March 31, 2026): purpose-built agents that run continuously over content workflows for tasks "that don't require human judgment," configured in natural language, all within the existing governance/permissions model. Architecture = a Main Agent (operate the platform conversationally) plus Expert Agents (specific high-value ops). Reported traction: ~60 organizations using the Agentic CMS; a Hostelworld case study cut 134 pieces in two languages to one day (Kontent.ai; CMSWire; Morningstar, 2026).
Contentstack — announced Agent OS on Sept 9, 2025 at ContentCon Europe, declaring "content management is dead" and pitching an Agentic Experience Platform (AXP) for the "context economy": Content Cloud + Data Cloud + Agent OS. General availability was "coming soon" at announcement — i.e., the vision shipped as a keynote before the product. Their no-code Automation Hub is the more concrete, already-available piece (Contentstack press, 2025; composable.com).
Cosmic — a smaller AI-native headless CMS taking the most aggressive agent posture: four agent types — Content Agents (generate/bulk/migrate), Code Agents (connect to GitHub, branch, write/fix code, open PRs, deploy to Vercel), Computer Use Agents (browser automation), and Team Agents (live in Slack/WhatsApp/Telegram). Plus a Cosmic CLI that scaffolds full-stack apps wired to your content. Pricing: Builder $49/mo, Team $299/mo, Business $499/mo, +$29/bucket and +$29/user (Cosmic, 2026). Assessment: ambitious surface area for a small vendor; Code Agents shipping code to production is the boldest claim and the one to verify carefully before trusting in a real pipeline.
A useful caution: much of the "agentic CMS" discourse in 2026 is vendor-authored (Cosmic's own blog comparisons, vendor "first agentic CMS" claims). Treat self-comparisons skeptically; the verifiable signals are documented APIs (Sanity Agent Actions, MCP servers) and real usage metrics, not keynote slogans.
The visual web-builder camp (generative build)
Here AI generates design and pages, not just text. These tools blur the CMS/site-builder line.
Webflow — prompt-to-production
Webflow has shipped an unusually broad AI surface and, notably, included it on every plan including free Starter:
AI Site Builder — generate a full multi-page site with a scalable design system from a prompt; built on Flowkit, Webflow's modular CSS framework, so output is editable and consistent. Webflow claims 60,000+ sites published via the builder.
AI Assistant — conversational; modifies page designs, generates new sections from the existing design system, and now powers app structure/logic/UI generation.
AI Code Gen / AI code components (GA'd April 30, 2026) — describe an interactive component (pricing calculator, multi-step form, search filter, location finder, animation) and generate a production-grade React component on the component canvas, in-line with your design system (Webflow Updates, 2026).
SEO/AEO — sitewide audits plus AI-generated alt text, meta titles/descriptions, and schema markup, explicitly framed for answer-engine optimization as well as classic SEO.
Pricing: Starter free (2 pages, webflow.io subdomain); Site plans $14–$39/mo (annual); Workspace Core $19/mo with code export (Webflow, 2026). Assessment: the most complete generative-build story among CMS-adjacent tools, and the code-component generation is a genuine step beyond layout generation.
Framer — Wireframer + Workshop
At its I/O 2025 event Framer shipped a coherent AI suite:
Wireframer — natural-language generation of responsive page structure/layout, directly editable in the Framer editor; Framer claims ~50% faster than manual.
Workshop — "vibe-coding" generation of interactive components (tabs, sliders, cookie banners, 3D hover cards) that automatically match the site's existing style tokens (colors/fonts/layout), with customizable parameter controls.
Plus Vectors 2.0 and built-in A/B testing / advanced analytics (Framer Spring Event 2025; AIbase).
Assessment: Framer's AI is tightly scoped to design generation and interactivity within Framer, not content operations. It is a website builder with a CMS, not a CMS with a builder — but for marketing sites the generative loop is fast and polished.
Builder.io — Visual Copilot + Fusion
Builder.io bridges design tools and code:
Visual Copilot — AI Figma-to-code, using an in-house model fine-tuned on 2M+ data points to convert designs to React, Vue, Svelte, Angular, Qwik, Solid, or HTML, one click from the Figma plugin. Its standout is component mapping: matching Figma components to components that already exist in your repo so generated code reuses your design system rather than emitting generic markup.
Visual Copilot 2.0 / "Design to Interactive" — make Figma designs interactive with natural language using your actual code, data, and APIs, inside the Visual Editor.
Visual Copilot CLI — granular Figma-to-code from the terminal.
Fusion — the broader AI visual-development environment that "sits between design and dev," where exported designs come alive for real-time AI editing against live code/data/APIs (Builder.io, 2025).
Assessment: Builder.io is the strongest design-to-code bridge and the most framework-agnostic. Component mapping is the feature that separates it from generic codegen. It is less a content store than a visual-development layer that can plug into one.
Best design-to-code bridge; component mapping to your repo
Legend: Shipped = documented, GA, usable today; Partial = available but narrow; Limited = minimal/early; Roadmap/Announced = stated direction or pre-GA keynote.
(As of 2026-06-05 — first-party MCP is now near-universal.) Beyond the vendors above, Strapi shipped its native first-party MCP server on May 28, 2026 (Beta; requires Strapi v5.47.0+; built into core as a route on the existing HTTP server; free / self-hosted only at launch; CRUD + publish/unpublish with a new scoped-token system) — Verified from Strapi's own blog. Hygraph (MCP server in Early Access since ~Jan 2026, all plans) shipped an update to its packaged AI Agents on May 27, 2026. Payload's official MCP plugin (@payloadcms/plugin-mcp) became extensible (other plugins can inject custom MCP tools, ~May 28). Storyblok's MCP Server (launched March 2026; 155+ tools, execute_readonly/execute_destructive permission model) is now stable/current rather than new. The net: every major headless CMS in this chapter now ships a first-party MCP server; the new differentiation has moved to agent observability (Sanity), coding-agent skills packages (Contentful, Sanity), and permission models — not whether an MCP surface exists.
What's worth stealing
Patterns from this cohort that any CMS-builder should consider adopting:
Schema-aware generation with validation (Sanity Agent Actions) — never let AI emit free text into structured fields; validate output against the content model and refuse invalid writes.
Drafts-by-default, never auto-publish (Sanity's June 2025 change) — AI mutations should be reversible and human-gated by default.
First-party MCP server (Sanity, Payload) — expose your content to external agents as the integration surface, rather than building a closed chatbot.
RAG-ready content (Payload) — auto-chunk and auto-embed so the CMS doubles as a retrieval store for downstream AI.
Bring-your-own-model (Storyblok) — let editors pick the provider/model; avoids vendor lock and lets quality/cost be tuned per task.
Component mapping (Builder.io) — generated output should reuse the existing design system/components, not generic markup.
AEO-aware SEO generation (Webflow) — generate schema.org markup and structured metadata aimed at answer engines, not just classic SEO.
Metering transparency — Contentful's AI Consumption Units are a cautionary tale: meter AI clearly, but don't let "add-on creep" make total cost unpredictable.
Key Takeaways
Structured content is the moat. The platforms whose AI is genuinely useful (Sanity, Kontent.ai) are the ones whose AI is schema-aware and validates output against the content model — free-text generation alone is commoditized.
The agent, exposed over MCP, is the new battleground. Sanity and Payload ship first-party MCP servers; Sanity's reported 1.5M agent tool calls is the rare real-usage datapoint in a field of keynote claims. (As of 2026-06-05, first-party MCP is near-universal — Strapi shipped a native MCP server May 28, 2026, joining Sanity, Contentful, Storyblok, Payload, Hygraph; differentiation has moved to agent observability and coding-agent "skills" packages.)
Consolidation has begun.Salesforce announced a definitive agreement to acquire Contentful on June 1, 2026 (pending close ~Q3 FY2027); treat Contentful's roadmap and pricing as volatile and weigh acquisition-integration risk on new selections.
Watch the shipped-vs-announced gap. Contentstack's Agent OS launched as a keynote ("content management is dead") with GA "coming soon"; Cosmic's Code Agents shipping to production is bold and should be verified before trust. Sanity's Agent Actions and Webflow's Code Gen are concretely GA.
Safety defaults matter. Sanity's June 2025 switch to drafts-by-default (instead of mutating published docs) signals how raw early agentic features were — reversibility and human gating are the right defaults.
Two distinct camps. Structured/headless (Sanity, Contentful, Storyblok, Payload, Kontent.ai, Contentstack, Cosmic) compete on schema-aware actions and agents; visual builders (Webflow, Framer, Builder.io) compete on generative design and code.
Pricing is fragmenting. Contentful meters AI in consumption units and sells AI/Studio as add-ons; Webflow bundles AI into every plan including free — opposite philosophies with real cost implications.
Bring-your-own-model is a differentiator. Storyblok's choice of OpenAI/Gemini/Perplexity/Claude per task, and Payload's BYO-key approach, give control that fully provider-managed platforms don't.
Payload's RAG-ready positioning is genuinely distinct — it treats the CMS as the embedding/retrieval substrate for your AI apps, not as an editorial copilot vendor.
Key References
Sanity. Agent Actions / Agent Actions changelog. 2025. https://www.sanity.io/agent-actions — Five action types (Generate/Transform/Translate/Prompt/Patch); June 2025 drafts-by-default change.
Webflow. Introducing AI code components / AI Site Builder. Apr 2026. https://webflow.com/updates/ai-code-components — Code Gen GA Apr 30 2026; AI Site Builder on Flowkit; AEO-aware SEO; AI on every plan.