{for each entry in 'ai-connections' ordered by trust}
Bring your own AI. Three ways in, one permission model.
Connect OpenAI, Anthropic, Google, OpenRouter, a custom endpoint, or run Ollama locally for free. What changes between these three is whose permissions the AI inherits, never whether permissions apply.
01 / Personal connector · your own role, at sign-in
Your AI, mirroring you.
One connector URL for the whole site. Everyone on the team pastes the same address into Claude, ChatGPT or Cursor, then signs in with their own Birkly account and gets exactly their own permissions back.
02 / Service agent · fixed role, set by an admin
A named bot with a fixed job.
When you need a shared identity rather than a personal one, like a publishing assistant that always acts as editor. Admins choose the role, and which roles or users are allowed to connect to it at all.
03 / Built-in assistant · role attached to the connection
Already inside the admin.
Birkly’s own chat, plus helper buttons through the interface where they are useful. Nothing to install: bring an API key, or point it at Ollama on your own machine and pay nothing at all.
An AI in Birkly can do exactly what its role allows, nothing more. Birkly checks the role before anything happens, so the same request gets a different answer depending on who is asking.
Every AI gets the same request“Publish the spring sale post.”
Maria’s ClaudeeditorPublished
Sam’s ChatGPTauthorRefused: authors can’t publish
Translator bottranslatorRefused: may only edit translations
Built-in assistanteditor + approvalWaiting for your approval
{from 'birkly' get 'approvals'}
The AI is a contributor, not an author.
Turn on require approval and writes from the built-in assistant stop in a queue instead of going live. You read it, you edit it, you decide whether it ships.
01
Asked
Someone asks the assistant for a draft, a field change, a new entry, in chat or from a helper button in the admin.
02
Checked
The backend resolves the connection’s role to a permission set and refuses anything outside it, before the action is ever queued.
03
Queued
Allowed writes become pending actions in the approval queue rather than changes to your live content.
04
Yours
You approve, edit or reject. Every decision lands in the activity log with a name attached.
External MCP clients confirm their own write tools in your AI app, so the approval step is never skipped. It just happens where you are working. In the docs: AI approvals
{for each entry in 'ai' where 'useful'}
More than a chat window.
AI shows up where it saves time, and always inside the same permission model.
Build the site with AI
Connect Cursor, Claude or ChatGPT and let it edit your site’s HTML, CSS and JavaScript together with the content, in one prompt.
Free and local
Point the built-in assistant at Ollama on your own machine. No API key, no cost, no data leaving the computer.
Helper buttons
Small AI actions where they help in the admin, such as drafting a description, without opening the chat.
No secrets for agents
AI can check the health of email and other connections, but cannot read credentials or change them without a human.