Build a content model from a prompt
Collections, singletons and blocks, created at runtime. No config file, no deploy.
Recipe · collection
Title
title
Cover photo
cover
Ingredients
ingredients
Steps
steps
Live in the API the moment it is saved.
The agentic headless CMS · Early access
Run your CMS from a chat
Your AI agent creates content models, writes, uploads media and publishes through MCP. No code, no deploys. Your team decides what goes live.
New chat · HelmCMS connectedMCP
Set up a recipes section for the Bakery site: title, cover photo, intro, ingredients and steps. Then add the overnight bagels recipe from my notes and publish it.
list_sitesBakery found
create_content_typeRecipe · 6 fields
import_media_from_urlcover photo
create_entrydraft · v1
publish_entrypublished
Done. The Recipe model is live in the API and “Overnight bagels” is published. Your site will refresh through its webhook.
Add a “Total time” field and fill it for every recipe.
update_content_type+ Total time (number)
update_entry ×6drafts saved
// Next.js: list published recipes
const res = await fetch(
`${process.env.HELM_API_URL}/delivery/v1/entries/recipe?limit=10&depth=1`,
{
headers: { Authorization: `Bearer ${process.env.HELM_LIVE_KEY}` },
next: { tags: ["cms"] },
},
);
const { data, nextCursor } = await res.json();{
"data": [
{
"id": "9f8cfa9c-…",
"slug": "overnight-bagels",
"type": "recipe",
"publishedAt": "2026-10-29T19:10:08Z",
"data": {
"title": "Overnight bagels",
"cover": { "url": "…", "width": 1600, "height": 1067, "alt": "…" },
"ingredients": [{ "item": "Flour" }, { "item": "Water" }]
}
}
],
"nextCursor": null
}Admin screens are real HelmCMS screens with demo content. The conversation and video are illustrations made from them.
One MCP server. Every assistant and framework.
- Claude
- ChatGPT
- Cursor
- Next.js
- Astro
- Nuxt
- SvelteKit
- React
Any
MCP client with OAuth
Claude, ChatGPT, Cursor
42
MCP tools your agent can use
Read, write, publish
0
Deploys to change a model
Models are data
16
Field types
Rich text to blocks
5
Roles on every level
Agents draft, editors publish
1
Request to read content
Next.js, Astro, Nuxt
Any
MCP client with OAuth
Claude, ChatGPT, Cursor
42
MCP tools your agent can use
Read, write, publish
0
Deploys to change a model
Models are data
16
Field types
Rich text to blocks
5
Roles on every level
Agents draft, editors publish
1
Request to read content
Next.js, Astro, Nuxt
“Add an FAQ block to every service page.”
Blocks
“Create a Recipe model with ingredients and steps.”
Content models
“Import these 12 photos and set alt text.”
Media
“Which pages are still drafts on the bakery site?”
Search
“Restore yesterday’s version of the pricing page.”
Versions
“Set up a new site for Harbor Clinic.”
Multi-site
“Add an FAQ block to every service page.”
Blocks
“Create a Recipe model with ingredients and steps.”
Content models
“Import these 12 photos and set alt text.”
Media
“Which pages are still drafts on the bakery site?”
Search
“Restore yesterday’s version of the pricing page.”
Versions
“Set up a new site for Harbor Clinic.”
Multi-site
Built for
Content work without tickets
Editors stop waiting on developers, developers stop shipping copy changes, and agents do the repetitive work.
Editors
Write, review and publish in a clean admin with drafts, versions and an Info panel that shows who changed what.
Developers
Model content once, read it from any framework with one REST call, and get signed webhooks when something changes.
Agencies
Every client in its own workspace with its own sites, members, keys and roles. One login for all of them.
AI agents
Give your agent a goal, not a checklist
Describe the outcome. The agent picks the tools: it builds the model, writes the entries, brings in media and publishes when you let it.
New chat · HelmCMS connectedMCP
Set up a recipes section for the Bakery site: title, cover photo, intro, ingredients and steps. Then add the overnight bagels recipe from my notes and publish it.
list_sitesBakery found
create_content_typeRecipe · 6 fields
import_media_from_urlcover photo
create_entrydraft · v1
publish_entrypublished
Done. The Recipe model is live in the API and “Overnight bagels” is published. Your site will refresh through its webhook.
Add a “Total time” field and fill it for every recipe.
update_content_type+ Total time (number)
update_entry ×6drafts saved
Design models
Create collections, singletons and blocks, add fields and reorder them, all over MCP.
Write and update
Draft new entries or update existing ones. Every save is a version you can restore.
Bring in media
Upload files or import from a public URL straight into the site's media library.
Publish with care
Publishing is its own permission. Keep it for people, or grant it when you trust the flow.
What your agent can do
Every part of the CMS. One conversation.
Text · required
Media
Array
Rich text
One agent, every client site
Organizations hold workspaces, workspaces hold sites. The agent only reaches what its user can.
Northwind Studio
bakeryandco.com
Bakery & Co
Bakery app
Bakery & Co
harborclinic.com
Harbor Clinic
Bring in media from anywhere
Direct uploads to R2 or imports from a public URL, with alt text, sizes and a searchable library.
Media library · bakeryandco.com
sourdough-focaccia.jpg
1600 × 1067 · 412 KB
cardamom-buns.webp
1600 × 1200 · 238 KB
menu-autumn.pdf
2 pages · 1.1 MB
Your site refreshes itself
Signed webhooks on publish, unpublish and delete, with retries and a delivery log.
Deliveries · https://bakeryandco.com/api/cms
entry.published
Overnight bagels
entry.unpublished
Rye crispbread
entry.published
Lemon olive oil cake
Every save is a version
Immutable history for every entry. Restore any earlier version as a new draft.
Overnight bagels · history
Version 4
Agent · 2 min ago
Version 3
Mira Okafor · yesterday
Version 2
Agent · Oct 27
Nothing goes live without an editor
Writing and publishing are separate permissions. Agents draft; people with the editor role publish.
Waiting for an editor
Brown butter cookies
Drafted by your agent
Rye crispbread
Drafted by your agent
PublishKeep as draft
Why HelmCMS
Other CMSs let agents edit content. HelmCMS lets them run the CMS.
What the agent can doCMS with code-defined modelsHelmCMS
Write and update entriesYesYes
Upload mediaYesYes
Create a new content modelNeeds code + deployYes
Add or change fieldsNeeds code + deployYes
Create a new site or client workspaceNeeds code + deployYes
The middle column describes CMSs whose schema lives in a config file. Many popular headless CMSs work that way; check your own setup.
Connect
Paste one URL into your assistant
The Connect AI agent page gives you the MCP URL and the steps. Your assistant opens a HelmCMS sign-in, you approve what it may do, and it starts working.
Scoped sign-in
OAuth 2.1 with read, write, publish and admin scopes.
One command
Claude Code connects with a single claude mcp add.
Model
Content models in seconds, no deploy
Models are data. Add a field in the builder or ask your agent; the API serves it right away. 16 field types, reusable blocks, sortable and tree collections.
Blocks
Page-builder blocks, nested up to three levels.
Menus
Tree collections with drag and drop for navigation.
Publish
Agents draft. People publish.
Give the agent the author role and it can write drafts all day. An editor reviews and decides what goes live.
Version history
Every save is an immutable version you can restore.
Audit log
Who wrote or published what, and whether it worked.
Multi-site
One CMS for every site you run
Organizations hold workspaces, workspaces hold sites. Each site has its own models, media, API keys and webhooks, and another client's content is not just hidden: it is reported as not found.
Northwind StudioOrganization · 2 owners
Bakery & CoWorkspace · client
bakeryandco.comAgent · author
Bakery appSite · 2 editors
Harbor ClinicWorkspace · client
Developers
Less glue code. One request.
A read-only delivery API with live and preview keys, filters, sorting, paging and related entries. Signed webhooks with retries tell your site when to refresh.
- Cached for 60 seconds on live keys, never on preview keys
- Per-site CORS origins for browser calls
- Keys shown once and stored only as hashes
// Next.js: list published recipes
const res = await fetch(
`${process.env.HELM_API_URL}/delivery/v1/entries/recipe?limit=10&depth=1`,
{
headers: { Authorization: `Bearer ${process.env.HELM_LIVE_KEY}` },
next: { tags: ["cms"] },
},
);
const { data, nextCursor } = await res.json();{
"data": [
{
"id": "9f8cfa9c-…",
"slug": "overnight-bagels",
"type": "recipe",
"publishedAt": "2026-10-29T19:10:08Z",
"data": {
"title": "Overnight bagels",
"cover": { "url": "…", "width": 1600, "height": 1067, "alt": "…" },
"ingredients": [{ "item": "Flour" }, { "item": "Water" }]
}
}
],
"nextCursor": null
}// Verify a HelmCMS webhook, then refresh
const ts = req.headers.get("X-CMS-Timestamp");
const sig = req.headers.get("X-CMS-Signature"); // "sha256=<hex>"
const expected = "sha256=" + hmacSha256Hex(SECRET, `${ts}.${rawBody}`);
if (sig !== expected) return new Response("bad signature", { status: 401 });
revalidateTag("cms");# Connect Claude Code to HelmCMS
claude mcp add --transport http helm https://api.your-domain.com/mcp
# Then just ask:
> Add a "Total time" field to Recipe and fill it for every recipe.Security
Hand the keys to an agent. Keep the control.
Scoped tokens
An agent's token never exceeds its user's role, and scopes cap it further.
Five roles, three levels
Viewer, author, editor, admin and owner on organization, workspace or site.
Immutable versions
Restore any earlier version of an entry as a new draft.
Audit log
Every write and publish with actor, action and outcome.
Safe imports
Media imports refuse private and local network addresses.
1,121 tests
Automated tests across contracts, services, database, server and admin.
Guides
Learn how it works
What is an agentic CMS?
An agentic CMS lets AI agents design content models, write, update and publish through the same API and permissions people use. What it needs and how HelmCMS does it.
Read the guide →
What is a headless CMS?
A headless CMS stores and manages content and delivers it through an API, so any website or app can display it. How it differs from a traditional CMS and when to use one.
Read the guide →
Connect a headless CMS to any frontend
How a frontend reads content from HelmCMS: the delivery API, API keys, draft previews and webhooks for revalidation, with Next.js and Astro examples.
Read the guide →
A CMS for agencies that run many client sites
Organizations, workspaces and sites with roles at every level: how HelmCMS keeps client content separate, lets AI agents draft safely and leaves publishing to people.
Read the guide →
Frequently asked questions
How is HelmCMS different from other CMSs that support AI agents?
In most of them an agent can read and write content, but the content model is defined in code, so a new field means a code change and a deploy. In HelmCMS the model is data: the agent can create and change models over MCP as well, so it can set up a whole site from a conversation.
Which AI assistants work with HelmCMS?
Any MCP client that supports remote servers with OAuth sign-in, for example Claude, ChatGPT and Cursor. Claude Code connects with one command. Clients that only accept a static token are not supported yet.
Can an agent publish without me?
Only if you allow it. Writing drafts and publishing are separate permissions. Sign the agent in as an author and it can draft but never change the live site.
Which frameworks can read the content?
Any of them. The delivery API is plain REST with a bearer key, so Next.js, Astro, Nuxt, SvelteKit, mobile apps and static site generators all work the same way.
What does it cost?
Pricing is not public yet. You can create an account and start today; if you run many sites, write to us and we will talk it through.
Where does it run?
The API, MCP server and delivery API run on Cloudflare Workers with D1 for data and R2 for media. The admin is a Next.js app.
Put your agent at the helm
Create an account, connect your AI agent and build your first content model in minutes.
Get started