Inkwire: Let Claude Edit Any Website, Not Just WordPress
WordPress now has an official way for AI agents to work on a site: the MCP Adapter, part of the AI Building Blocks for WordPress initiative. Ask Claude to fix a page and it can, as long as the site runs WordPress.
Most of the sites we look at in DadSEO don't. They are Next.js marketing sites, Astro docs, an Express app with a blog bolted on, a FastAPI backend that serves a few landing pages. For those there was no equivalent, so we built one and opened it up: Inkwire, a small open protocol that lets Claude, Cursor or any MCP agent read and edit a website's content.
This post was written, checked and published through it. Further down is what Inkwire said about the first draft.
What Inkwire is
Inkwire is a REST + JSON contract between a website (the receiver) and whoever writes to it (a person's agent, a publishing tool, a script). It is MIT-licensed and deliberately small: the only required endpoint is POST /api/content. Everything else is a feature the site advertises at GET /api/inkwire, so a client knows what a site can do before it tries.
What a site can offer:
- Posts, pages and other kinds of entries, written in Markdown that the site renders to sanitized HTML.
- Upsert by
external_id: send the same id again and the entry updates in place instead of duplicating. - List, read, edit and delete, with
ETag/If-Matchso two editors never silently overwrite each other. - Tags, categories and scheduling (a published entry with a future date waits until then).
- A media library: upload, browse and set alt text.
- Full SEO control: title, description, Open Graph, JSON-LD, canonical and noindex, with sensible values derived from the entry when you set nothing.
- Undo, SEO warnings, redirects and translations, covered below.
There are reference receivers for Next.js, Express and FastAPI, clients for Node and Python, and a conformance suite you can point at any site, in any language, to prove it speaks the protocol.
Set it up in two commands
On a Next.js or Express site, one command writes the routes, adds the dependency and generates two keys (one full, one read-only):
npx create-inkwire
Then connect your agent. In Claude Code:
claude mcp add my-site --env INKWIRE_KEY=<key> -- npx -y inkwire-mcp https://my-site.com
Claude Desktop and Cursor take the same command in their MCP config. From there you ask in plain words: "Our plans now start at $12: update the pricing page and its meta description. Show me the change first."
You don't need a database. A receiver can keep entries as files in the repo, commit them to GitHub (useful on Vercel, where the disk is read-only), or sit on top of the database you already have. Our own blog is the last case: the receiver writes straight into the Postgres table the site already reads.
What the agent can do
The MCP server reads the site's discovery at startup and only offers tools the site supports. A site without delete has no delete tool.
| Tool | What it does |
|---|---|
list, get | Find entries by kind, status, words, tag or language, and read one with the SEO values the site really uses |
create | A new post or page, as a draft unless told to publish |
update | Exact find-and-replace edits on the text, or any field |
history, restore | Earlier versions of an entry, and putting one back |
terms | The tags and categories already in use |
upload_image, list_images, update_image | The media library and its alt text |
Built so you can hand an agent the keys
Giving an AI write access to a live site is a reasonable thing to be nervous about. The protocol answers that with mechanisms, not promises:
- Preview, then apply. Every create and update can run with
dry_run=true, which returns the result (and its warnings) without writing anything. The agent is told to show you the before and after and to wait for a yes. - Undo. Every update, restore and delete first keeps the version it replaces. "Undo that" is one restore, a deleted entry included, and a restore can itself be undone.
- No silent overwrites. An edit carries the
ETagof the version the agent read. If someone changed the page in between, the write is refused and the agent reads it again. - Exact edits. Each piece of text to replace must occur exactly once on the page, or nothing changes.
- Drafts first. New entries are drafts unless the agent is told otherwise.
- Read-only keys. Give an agent the read-only key and it can audit every page and change none of them.
SEO checks on every write
This is the part we care about most at DadSEO. Every write, and every dry run, comes back with a list of warnings from the site itself:
- a title that is too long, or that another page already uses;
- a missing meta description;
- a second H1, or a heading level skipped;
- an image without alt text;
- a link to a page that doesn't exist, is still a draft, or has moved;
- link text like "click here" or "read more".
Warnings never block a write. They tell the agent what to fix before it publishes, which is the cheapest moment to fix anything.
Three more things happen without anyone asking:
- Renames keep their rankings. Change a slug and the old address answers with a permanent (308) redirect to the new one, so links and search signals follow the page.
- Translations get hreflang. Entries that share a
translation_keyare linked, and each page lists every language version, itself included. - Images don't shift the layout. Uploads record their width and height, and library images get both in the HTML.
What it said about this post
We sent the first draft of this article as a dry run. The site answered in a quarter of a second, wrote nothing, and returned two warnings:
[
{ "code": "description_missing", "field": "seo.description",
"message": "no description: search results show the title twice; write an excerpt or seo.description" },
{ "code": "link_broken", "field": "markdown",
"message": "links to pages that are not on the site: /blog/we-audited-our-own-ai-visibility" }
]
The first was right: the draft had no excerpt, so search results would have shown the title twice. We wrote one.
The second was a false alarm: that post is a Markdown file in our repository, not an entry the receiver manages, so the check could not see it. The site can now tell the link check about pages like that.
WordPress MCP Adapter or Inkwire?
They solve the same problem for different sites, so the choice is mostly made for you.
| WordPress MCP Adapter | Inkwire | |
|---|---|---|
| Runs on | WordPress | Anything that can serve HTTP: Next.js, Express, FastAPI, Astro, your own stack |
| What the agent gets | The abilities plugins and themes register (private until marked public) | Content: posts, pages, media, SEO fields |
| Content model | WordPress's own | Markdown, rendered and sanitized by the site |
| Undo, dry runs, SEO warnings | Up to each ability | Part of the protocol |
| License | GPL-2.0-or-later | MIT |
If your site is WordPress, use the adapter. If it isn't, Inkwire gives your agent the same reach in about two minutes.
Who uses it
- DadSEO's blog, the one you're reading, runs an Inkwire receiver over Postgres, and passes the conformance suite there.
- Anaella connects to any Inkwire site with its address and key, and its agent reads and edits that site's posts, pages, SEO fields and media the way it does a WordPress site's.
If you're checking whether AI tools can find and cite your pages at all, that is a different problem. We wrote about it in our own AI visibility audit.
FAQ
Is Inkwire free?
Yes. The protocol, the receivers, the clients and the MCP server are open source under the MIT license, on GitHub and npm (Python packages on PyPI).
Does it work with Claude Desktop, Cursor and other agents?
Any MCP client that starts stdio servers works with inkwire-mcp. Without MCP, the Node and Python clients, or plain HTTP, reach the same API.
Do I need a database?
No. Entries can live as Markdown files in the repo, in a GitHub repository, or in the database you already run.
Can an agent break my site with it?
It can only touch content, never code or settings. With the read-only key it can't write at all. With a full key, every change can be previewed first and undone afterwards.
Founder of DadSEO. I build tools that turn SEO data into strategy — not scores. Previously spent years running audits that told me what was broken without telling me what mattered.
Read more about me →Ready to see what your SEO data actually means?
DadSEO connects to Google Search Console and gives you a strategic diagnosis — threats, gaps, and opportunities ranked by impact.
Get Your Free Audit