The MCP connector
Verand can be added to Claude, ChatGPT, Cursor, VS Code and other AI clients as an MCP connector. Once it's connected, you can ask your assistant about your site and it reads the answer out of Verand directly, and you can ask it to start work that would otherwise mean opening the dashboard.
MCP is the Model Context Protocol, an open standard for connecting an AI client to a tool it can use. Verand runs an MCP server, your client connects to it, and everything it does happens as you, with your permissions, on your plan.
The connector cannot publish. Not by policy, by construction: there is no publish tool for it to call. The same is true of overriding a compliance block and of deleting anything. More on that below, because it's the part worth understanding before you connect.
📸 Screenshot slot:mcp-ai-clients-card.png
Settings → Integrations, the AI clients card, with the endpoint and the client tabs visible.
Connecting
You'll find everything you need at Settings → Integrations, in the AI clients card. It shows your endpoint and step-by-step instructions for each client.
There is no API key. Verand doesn't issue one. Connecting opens a sign-in and a consent screen, which is the whole authorization step.
Claude
On the free Claude plan you can have one custom connector. Pro, Max, Team and Enterprise have no such limit. On Team and Enterprise an Owner may need to add it for the workspace first.
- Open Customize → Connectors
- Click +, then Add custom connector
- Paste your Verand endpoint
- Sign in with your Verand account and approve the access it asks for
ChatGPT
ChatGPT needs a paid plan, and the web app. Developer mode is on Plus, Pro, Business, Enterprise and Education, on the web only. Not on the free plan. On Business and Enterprise an admin can switch it off for everyone.
- Open Settings → Security and login and turn on Developer mode
- Go to connectors, click +, and create an app for your MCP server
- Paste your endpoint and set authentication to OAuth
- Sign in and approve
Claude Code
Run this in your terminal, then run /mcp inside Claude Code and choose to authenticate (or claude mcp login verand from your shell):
`` claude mcp add --transport http verand https://app.verand.ai/mcp ``
Cursor and VS Code
Add this to your MCP config file, which is ~/.cursor/mcp.json for Cursor:
``json { "mcpServers": { "verand": { "url": "https://app.verand.ai/mcp" } } } ``
Reload the window. Your editor opens a browser to sign in.
Anything else
Any client that supports remote MCP servers over HTTP can connect. Paste the endpoint wherever it asks for a server URL, and choose OAuth if it asks how to authenticate.
The consent screen
Every route ends on a Verand consent screen, and it's worth reading rather than clicking through. Two things on it matter.
The client's name is self-reported, and the screen labels it that way. Any client can register itself with any name and any logo, so the name on the screen is a claim, not a verified fact. The redirect host is the part you can trust, because that's where the access actually goes. If you're connecting Claude and the redirect host isn't Anthropic's, stop.
Access is granted to your whole account, not to one site. Whatever the connector can do, it can do to any site you can reach in Verand.
What it can do
- Read your sites, content, site health and rankings
- Generate drafts and run analyses. These use your plan's monthly allowance, the same as work started from the dashboard
- Open a published page for updating, and add topics to your content plan
- Save things you tell it about your business to your knowledge base
- Move a draft between work-in-progress and ready-for-review
What it can never do
- Publish anything to your live site
- Override a compliance block
- Delete any of your content
That second list is not a promise about behaviour, and the distinction is the point. There is no publish tool, no compliance-override tool and no delete tool for a client to call, so no instruction, mistake or prompt injection can reach one. A model that decides to publish your draft finds nothing to invoke.
The one adjacent case: the connector can move a draft to ready for review. It cannot move it to approved or published. Approval stays where it has always been, with a person in the dashboard.
It spends your allowance
This is the part customers are most often surprised by, so it's stated plainly on the consent screen and it's worth repeating here.
Asking your assistant to write an article consumes an article from your monthly allowance, exactly as if you had clicked Generate in the dashboard. Opening a page for refresh consumes a content refresh. Running a routine draws on your on-demand allowance, and a few routines cost real money in provider fees. Reading anything is free.
Metering is not separate for MCP. The connector calls the same code the dashboard buttons call, so the ceiling, the counting and the caps are identical. Ask your assistant to check your usage before a long session and it will read the same numbers your Plan and Usage tile shows.
A model can start work in a loop. Rate limits cap how fast that can happen, but your plan allowance is the only ceiling on how much of it gets spent. If you're handing an agent a long task, ask it to check usage first.
What your assistant can see
The connector exposes 23 tools, grouped by what they do:
| Group | What your assistant can ask for |
|---|---|
| Orientation | Your sites, a site overview, your plan usage and what's left |
| Content | Your articles, one article with its compliance gate result, a semantic search across your published pages |
| Site intelligence | Site issues by severity or category, one crawled page with its scores and Core Web Vitals, your opportunities |
| Market data | Keyword rankings with competitor positions, AI search visibility per engine, competitor authority, Search Console and Analytics by page |
| Research | Your briefs, keyword metrics, and a proposed fix for a failing check |
| Writes | Generate a draft, check a run, move a draft's status, run a routine, open a page for refresh, add topics to the plan, save experience facts |
Two of those deserve a note.
Proposing a fix writes nothing. It returns the suggested replacement plus the before-and-after validation, so you can see whether it would actually clear the check. Applying it is a separate action you take in the editor. Compliance and critical checks are refused outright.
Saving experience facts is the one place the connector writes your own words into Verand. If you tell your assistant something true about your business during a conversation, it can save that to your knowledge base, where it feeds future articles. That's useful, and it's also why the consent screen names it.
Which plans include it
All of them. The connector isn't tier-gated. Your plan allowance already governs what any of it costs, so gating the door as well would charge twice for one thing.
This is separate from the API access listed on the Agency plan. API access is for agencies building on top of Verand. The MCP connector is for using your own AI client. They are different things.
Revoking access
Settings → Integrations lists every client you've authorized and lets you revoke any of them.
One honest detail about timing: revoking blocks the client immediately at Verand's end, but an access token that was already issued can otherwise stay valid for up to about an hour with the identity provider. Verand checks its own revocation list on every request specifically so that gap doesn't exist in practice, and that check fails closed. If you revoke because something is wrong, the connection stops working straight away.
📸 Screenshot slot:mcp-authorized-clients.png
The authorized clients list with at least one connected client and its revoke control.
Troubleshooting
"It says it can't publish my article." That's correct and deliberate. Approve and publish in the dashboard.
The client connects but every request fails. The most common cause is signing in with an account that isn't a member of the organization holding the site. Check which account the consent screen used.
My assistant says a site doesn't exist. Ask it to list your sites first. It works from the ids it gets back, and a name it guessed will not resolve.
Nothing happens after I paste the URL. Check the client's own requirements above. On ChatGPT the most common cause is Developer mode being off, or being on a plan or a platform that doesn't have it.
Work started from my assistant is missing. It isn't lost. Article generation takes two to four minutes, and everything the connector starts appears in your Runs feed and in the dashboard exactly as if you'd started it there.