Ship content without a marketing hire.

The article you are reading was drafted by an agent on a schedule — a recurring task on an ordinary Claude subscription, pointed at Munin over MCP — then filed as a draft and left for me. That is not a demo. It is how this Journal has been produced for weeks, and the interesting part is not that a language model can write, because everybody knows that by now. It is that the thing doing the writing and the thing holding the tools are two separate purchases, and only one of them is ours.

Can you run content marketing without hiring a marketer?

Yes, for the production half. An agent can pick a topic off a rotation, verify its claims against live sources, draft into a typed CMS schema, attach a cover image, cross-link the existing archive, and file the result as a draft — on a cadence, with nobody watching. What it cannot do is decide what your company should be arguing. That decision stays with a person, and it is the part actually worth their week.

What does the rig actually consist of?

Four things. A CMS collection whose article body is an ordered array of typed blocks. A handful of markdown skills describing the editorial job. A scheduled runner holding an MCP connection. And a review queue, which in practice is the word draft. Munin publishes 204 MCP tools and 47 bundled skills at one endpoint; the content job touches maybe fifteen of them.

None of that is a pipeline somebody built. It is four calls in a loop, and the loop is short enough to read in one sitting.

• 01Read the shelf first. cms_list_entries on the collection, then a look at the last five topics. The rotation is written into the skill, not into the model's judgment: three of every five pieces have to attach to search demand that already exists, at most two can be house argument. An agent will happily write essays forever if you let it.
• 02Verify, then draft. Every external claim — a competitor's price, a licence, a date — gets checked against the vendor's own page before it goes in. Then cms_create_entry with status: "draft", a body of typed blocks, and a slug shaped like the query it targets.
• 03Dress it. The runner generates a cover with whatever image model it has, cms_upload_asset_from_url puts the file in the media library, and cms_update_entry attaches the returned asset id to the entry's cover_image field. Every update carries ifVersion, so two agents editing the same entry produces a conflict rather than a silent overwrite.
• 04Curate the back catalogue. The same pass that writes one new piece re-reads an old one: prices that have moved, a link that has died, an article missing the two internal links that would have made it useful. skill://cms/review-stale-entries is a curator pass by design — it reports and edits, it does not delete.

Who supplies the intelligence?

You do, and that is the design. Munin ships the tools; the model arrives with whatever MCP client you point at the endpoint. Connect a Claude subscription, a ChatGPT workspace, Cursor, Codex, Gemini or a runner you wrote yourself, and the writing is done by a frontier model you chose and can change. There is no Munin model, because the intelligence was never ours to sell.

That sounds like an architecture note and it is really a pricing one. Take the setup behind this Journal. Claude's scheduled tasks are available on every paid plan — Pro at $20 a month or $200 a year, Max at $100 or $200 a month, plus Team and Enterprise — and you give a task a prompt, an approval mode, a cadence of hourly, daily, weekly, weekdays or manual, and, optionally, the model you want it to use. They run remotely, so the run fires on Friday morning whether or not the laptop is open. Point that at https://mcp.getmunin.com and the whole editorial loop above is a subscription line item you may already be paying for, not a per-article charge.

Be clear-eyed about the ceiling: a consumer subscription has usage limits, and at real volume you move to API billing and pay tokens. But tokens are a commodity with four vendors competing to sell them to you, and the number on that bill is not set by the company that also sells you the CMS.

The part that compounds is the upgrade path. Models get better roughly every quarter, and in this arrangement adopting a better one is editing a field on a scheduled task. Nobody has to route to it on your behalf, nothing needs to appear on a roadmap, and the tools underneath do not change — the same cms_create_entry call is being made, just by something that writes better than it did in June.

It also means the client is a preference rather than a commitment. The same 204 tools answer a terminal, a desktop app and a web workspace with the same scopes and the same audit log, so a developer and an operations person can work the same content collection from completely different surfaces. The ninety-second version, using Claude Desktop and one HTTPS URL, is written up separately.

Who decides what goes live?

You do, and the mechanism is boring on purpose. Every entry the agent creates lands as a draft. Publishing is a different tool — cms_publish_entry — and an agent holding a key without that scope cannot reach it. Entries are versioned, so a bad edit is a rollback rather than an incident, and every tool call is audited with a name attached.

This is the same shape as every other automated pass in the platform: outreach drafts wait in a review queue and a support answer becomes a knowledge base article only once a human promotes it. The queue is not a training-wheels phase. It is where the judgment lives, and it is the reason you can leave the rest running unattended.

What does HubSpot Content Hub cost to do this?

Content Hub Professional is $450 a month on an annual commitment, $500 paid monthly, including three Core Seats; additional Core Seats start at $45 a month. Enterprise is $1,500 a month with five seats, additional seats from $75. Professional includes 3,000 HubSpot Credits a month, Enterprise 5,000, and further credits are $9.00 per thousand on annual billing or $0.010 each. All of that is on HubSpot's own content pricing page, read on 21 August 2026.

And it buys a great deal. A drag-and-drop page builder a non-technical marketer can actually use, landing pages, forms, A/B testing, SEO recommendations tied to the CRM, ad attribution, and a support organisation behind all of it. If your content operation is four people who need to ship campaign landing pages this quarter without waiting on engineering, that is money well spent and I would not try to talk you out of it.

The difference that matters for this rig is who owns the model. HubSpot publishes model cards, which is more disclosure than most vendors bother with and worth crediting properly: the Breeze bucket its Content Hub features sit in lists OpenAI's GPT-4.1 family, GPT Image-1 and Anthropic's Claude Sonnet 4 among others, and the Agents bucket lists Claude Opus 4.5. Those are real models doing real work, with zero data retention and no customer data used for third-party training. But the list is theirs. You do not pick from it, you cannot substitute your own, and the output is denominated in HubSpot Credits — a per-action price set by your vendor, on a rate sheet your vendor revises.

Munin's side of that is short, because there is nothing to describe. No Munin model, no Munin credit, no rate sheet. Cloud Free is €0 a month with 5,000 MCP calls, 250 contacts and 100 MB of assets, which is enough to run a journal at this cadence, and Cloud Premium is a €99 flat monthly price per organisation with metered overage, launching soon. The model bill goes to whoever you already pay for a model, at whatever they charge — and it does not move when you publish twice as often.

What if what you want is just a very good blog?

Then buy Ghost, and enjoy it. Ghost(Pro) is $18 a month for Starter, $29 for Publisher, $199 for Business on yearly billing, per Ghost's pricing page as of 21 August 2026. The editor is a pleasure, the newsletter and paid-subscription machinery is built in and takes no transaction fee of its own, the Content API is on every plan and the Admin API from Publisher up, and the whole thing is MIT-licensed and self-hostable if you would rather run it. For a publication — one that is the product rather than an argument attached to one — it is a better fit than a customer platform, and there is no version of this comparison where I pretend otherwise.

Where the two diverge is what happens after publication. Ghost knows about members and subscriptions. It does not know that the person who read your architecture piece on Tuesday opened a support conversation on Thursday and became a deal on the following Monday, because those three facts live in three products with three different ideas of what a person is. In Munin the CMS is one of six modules over a single Postgres with row-level security carrying the tenancy, so the reader, the thread and the deal are rows against one contact record. That is the whole architectural bet, and content is where it pays off first, because content is the thing you publish before anyone has told you who they are.

How do you tell whether any of it worked?

By asking which article produced a customer, which is a query rather than a project. Every CMS delivery response already ships a _tracking block keyed by the stable entry id, so per-entry views are recorded without minting a key or dropping a third-party script. Those view events sit in the same database as the contact record, so an anonymous reader who later identifies themselves carries their reading history onto their own timeline.

That turns the useful question round the right way. Not which post got traffic — traffic is cheap and an agent will happily manufacture it — but which post was read by the people who eventually bought something. We wrote up how that attribution actually resolves, and it is the number that should decide the next topic on the rotation. Feed it back in and the rotation stops being a guess.

What is not in the box

Munin ships no visual page builder, no drag-and-drop landing page designer, no A/B testing, no ad-campaign manager and no keyword research tool. Content here is a typed collection, a delivery API and a tool catalogue. If your plan requires a marketer to assemble a landing page in a canvas this afternoon, that is a real requirement and the platforms above meet it.

What this rig is for is the other job — the one where the constraint is that nobody has two clear hours to write, and the archive rots because nobody owns it. That job is almost entirely mechanical once the argument is decided: verify the claims, hold the house voice, keep the internal links current, notice that a competitor changed their price in July. Those are exactly the tasks that survive being handed to a scheduled agent, and they are the tasks that never get done otherwise.

So the honest scoping question is not which tool is better. It is whether your bottleneck is assembly or authorship. If it is assembly, buy the canvas. If it is authorship, a rotation, a typed schema and a review queue will out-produce a marketing hire you have not made — written by the best model on the market this month, because that is a setting rather than a supplier decision, and measured against a contact record that tells you afterwards which piece was worth writing. Solo founders hit this wall first, which is why the free tier is sized for exactly this.

Frequently asked questions

Which AI model writes the articles? Whichever one you point at Munin. Munin exposes 204 MCP tools and no model of its own, so the intelligence comes from your client — Claude, ChatGPT, Cursor, Codex, Gemini or a custom runner. On a Claude scheduled task the model is an explicit field on the task, so switching to a newer one is an edit rather than a migration.

Can AI write my blog posts and publish them automatically? It can write and file them. Publishing should stay a human call, and in Munin it is a separate tool an agent can be denied. Every entry is created as a draft and versioned, so approval is one action and reversal is a rollback. Fully unattended publishing is technically available and a bad idea for anything with your name on it.

How much does it cost to run content marketing with an AI agent? Two bills, and neither is per article. Munin Cloud Free is €0 a month with 5,000 MCP calls, 250 contacts and 100 MB of assets, or self-host for the price of a small server. The model is your own subscription or API account — Claude Pro is $20 a month, Max $100 or $200 — charged by your provider, not by us.

Do I have to use Claude specifically? No. Munin serves standard MCP with no client-specific behaviour, so any MCP-capable client connects to the same endpoint with the same tools and the same audit log. Claude is what this Journal happens to run on because its scheduled tasks are on every paid plan and run remotely on a cadence.

Is an AI-written blog bad for SEO? Google's stated position is that it rewards helpful content regardless of how it was produced and penalises content made primarily to manipulate rankings. In practice the failure mode is not the model — it is publishing unverified claims at volume. The fix is editorial: verify every external number against the vendor's own page, and keep a human approval step so nothing ships unread.

What is the best CMS for an AI agent to write into? One with a typed content model the agent can fetch, per-block addressing so edits are patches rather than rewrites, optimistic concurrency on writes, and versioning. Sanity's Portable Text and Strapi's dynamic zones are strong models for the body itself. Munin adds the tool catalogue and the contact record the content is measured against.

Do I still need a marketer? For strategy, positioning and knowing which argument matters this quarter — yes, and that is where their time should have been going. What comes off their plate is drafting, formatting, cover images, internal linking and archive maintenance.

Can the agent update old articles too, or only write new ones? Both, and the second matters more over time. A curator pass re-reads published entries, corrects prices that have moved, fixes dead links, and adds internal links where they fit the sentence. Pass ifVersion on every update so a concurrent edit fails loudly instead of quietly winning.

The skills this runs on are markdown you can read before you trust them — they are in the repository, and Munin Cloud will hold a collection for you at no cost while you point your own model at it.

A marketing hire you never made can't write. A rotation, a schema and a model you already pay for can.