---
title: "AI agent access"
description: "For account owners: what a connected assistant may do, approvals, the off switch, the call log and revoking."
canonical: "https://writavo.com/docs/agents"
last-updated: "2026-09-29"
---

# AI agent access

When someone connects Claude, ChatGPT, Cursor, Codex or another AI assistant to your Site, it gets a limited key. Here is what it may do, what it must ask first, and how to cut it off.

## What a connection is

Connecting an assistant through the hosted or local MCP server creates an AI agent key for each Site it reaches, labelled with the app and limited to what the person allowed. Keys made by hand in the dashboard are not agent keys, and nothing here applies to them.

- **Sites:** pick a default Site and add others, or every Site in the organisation (including future ones). The assistant lists them with `list_sites` and names one with `site`; it cannot reach any other. Change them in Settings > AI agents > Sites on this connection, without signing the assistant out. Only the person who connected it can add a Site.
- **Never more than the approver.** If their role on the Site narrows, the connection narrows on its next request.
- Setup is in [the MCP server guide](/docs/mcp). What each area can do is in [Settings and administration](/docs/administration).

## Connecting an assistant

1. **Sign in.** The assistant opens a Writavo page. Sign in to app.writavo.com, or create an account. Nothing is pasted.
2. **Check who is asking.** The page shows the app's name and, for the hosted server, the return address. Continue only if you just started this yourself.
3. **Choose the Site.** Only Sites where you can manage API keys. Sites whose organisation turned AI agent access off cannot be chosen.
4. **Choose permissions and approve.** The picker starts from your organisation's default. The assistant connects a few seconds later.

## The permission picker

Each area is No access, Read, or Read and write. Site details (name, domain, organisation) are always readable.

| Area (what Read shows) | Read and write also allows |
|---|---|
| Articles (drafts included) | Create, edit, publish, schedule, unpublish, delete. With your permission: import a moved blog's views, reactions and shares |
| Content types and entries | Create, edit, publish, schedule, delete entries. With your permission: change content types and their fields |
| Categories, tags and authors | Create, edit, delete |
| Media library | Upload and delete images |
| AI pipeline (status, configuration, content plan) | Read, plan and run (spends credits): edit the content plan, start runs, send articles back for a rewrite. With your permission: change configuration, prompts and formats |
| Site settings and design (knowledge profile, hosted site branding) | Change them |
| Publishing and domains (how the blog is delivered) | Can cost money: set up the reverse proxy, connect or remove a custom domain, publish the hosted site. With your permission: disconnect and push to an external CMS |
| SEO suite | Spends credits: curate competitors, keywords and backlinks, run paid scans |
| Outreach contacts (readiness, contacts, campaigns, threads, with names and email addresses) | Read only |
| Reader comments (settings and queue, with commenters' email addresses) | Read and moderate: turn comments on or off, approve, reply, mark as spam, delete, erase a commenter's comments |
| Team and organisation (members, invites, permissions) | Invite, change roles, remove members. With your permission: change member and role permissions, rename the organisation |
| Billing (plan, balances, invoices, caps) | Start a credit pack purchase, change auto-refill and the Site's credit cap. Never card details or the plan |
| Reports and logs (overview, costs, visitor analytics and its install check, logs, the agent call log) | Read only |

- You can never grant more than you have. A choice you cannot grant is disabled, with the reason; "with your permission" parts are added only if you hold that permission.
- Reader comments needs your own Moderate comments permission (owners, admins and editors by default).
- **Read only** sets every row to Read, except Outreach contacts and Reader comments (left off unless you chose them).
- **Older connections** do not get newer permissions automatically: press Add (or Permissions) on the connection in Settings > AI agents, or Scopes in Settings > API keys. No reconnect needed.

## The default for new connections

- **Out of the box:** Read and write for articles, content types and entries, categories, tags and authors, media, and Site settings. Read for everything else, except Outreach contacts and Reader comments (off). So no spending money, delivery or team changes unless someone picks that row.
- An organisation that changed its default keeps its own.
- Change it in Settings > AI agents > Default permissions for new connections (needs permission to manage the organisation; owners and admins by default). It applies to every Site in the organisation, and existing connections keep what they have.

## Approvals

Some actions wait until a person says yes:

| Always asked | Asked while approvals are on (the default) |
|---|---|
| Requesting a pipeline run, turning the pipeline on or up, an AI rewrite or new images for an article | Deleting an article, category, tag, author or media file |
| The seven paid SEO scans | Unpublishing an article |
| Setting up the reverse proxy, connecting or removing a custom domain, publishing the hosted blog | |
| Disconnecting the external CMS, pushing articles to it | |
| Deleting an article format | |
| Turning on auto-refill, raising or removing the Site's credit cap, keeping the current plan | |
| Inviting someone, changing a role, a member's permissions or a role's permissions, removing a member | |

1. **The assistant gives you a link.** https://app.writavo.com/approvals/<id>. Nothing has happened yet. A link to any other address is not from Writavo.
2. **You approve or deny.** The page shows the action and its kind (Spends money, Changes the team, Changes the live site, Cannot be undone), the item, the Site, which assistant asked and when it lapses.
3. **The assistant retries.** After approval the same request goes through. After a denial it must not try again.

- **Who decides:** anyone on the Site who could take the action themselves. A team change can only be approved by the person whose assistant asked, never by the member it affects.
- **One approval, one request,** used once. If anything changes before the retry (the article, a price, a role), the assistant must ask again. Undecided requests lapse after 24 hours.
- Approvals apply to any MCP client, even one using a key made by hand. Pending ones are listed in Settings > AI agents.
- An owner or admin can turn approvals off there. That stops only the delete and unpublish questions; everything else is still asked, and permissions still apply.

## What an assistant can never do

Whatever its permissions, the server refuses these and the assistant points you to the right page:

- Approve its own requests, give itself more access, or change your role or permissions.
- Change the agent switch, approvals or the default permissions.
- Delete a Site or the organisation, transfer ownership or make anyone an owner.
- Enter card details or change the plan. It can give you the billing link, or a Stripe Checkout link for a credit pack you pay for yourself.
- Accept the outreach policy, pass the identity check or connect an outreach mailbox.
- Enter a WordPress, Ghost or Bing Webmaster Tools credential.
- Create, rotate or revoke API keys, or manage webhooks.

Full list, with the page for each: [Settings and administration](/docs/administration#never).

## Turning agent access off

- Settings > AI agents > Allow AI agents covers the whole organisation (needs permission to manage the organisation).
- **Off:** every agent key is refused from its next request (`403 AGENT_ACCESS_DISABLED`), reads included, and no new connection can be approved. An assistant can still sign out.
- **On again:** existing connections work at once, with no new sign-in. Keys made by hand are never affected.

## Seeing which apps hold a key

Settings > AI agents lists each connection by app (for example "Claude Code (hosted MCP)"), where it runs, its key prefix, permissions, dates and expiry, with Permissions and Revoke buttons. The same keys show in Settings > API keys with an AI agent badge.

## The call log

Settings > AI agents also logs every agent-key request, refusals included: time, app, MCP tool, operation, outcome and duration. Kept 90 days; needs permission to view logs.

## Revoking

- Revoke in Settings > AI agents or Settings > API keys. It stops on its next request and its sign-in is not renewed.
- Revoking on the Site it was approved on ends it everywhere; on another Site, it removes only that Site.
- To reconnect, connect the assistant again and approve a new key.

## How long a connection lasts

90 days, renewed while in use: a use within 30 days of expiry pushes it to 90 days from then, up to 365 days from creation, after which you connect again. It is not renewed if the approver can no longer manage API keys on the Site, or while agent access is off.

## The security model

- The hosted server keeps the key encrypted on Writavo's servers. The assistant only gets OAuth tokens for mcp.writavo.com (access tokens last one hour), never the key.
- Sign-in must start and finish in the same browser. A link someone else started is cancelled and any key it made is revoked.
- Every request is checked on the server against the key's permissions and the approver's live role.
- Approvals, denials, revocations and changes to these settings go in the organisation's audit log.

> A key you create by hand and paste into an assistant gets none of these controls. Connect through the MCP server instead.
