AI connector
The AI connector is a Model Context Protocol endpoint you point your own assistant at — Claude, or anything else that speaks MCP. It lets an assistant read your test library and write to it: create test cases, add steps, report bugs, run sessions, read dashboards.
:::info Plan The MCP server is available from the Starter plan upwards. :::
The security model, in one paragraph
The connector implements no authorization of its own, and that is the point. A tool call resolves your connector token to your user, mints a 60-second internal token, and makes an ordinary request to the same API the browser uses. Every tenancy check, every role rule and every policy applies exactly as it would to you in a browser, and anything the assistant creates is stamped with you as its author. A rule written only in the connector would be a rule the web app does not have.
Two consequences worth stating plainly:
- An assistant can never do something you could not do. If you are a member, it cannot delete.
- No delete tools exist at all, and the internal allow-list refuses DELETE on product data even if one were added. Deleting a folder cascades through its test cases and steps with no undo, and that is not a thing to hand to a model.
Creating a connection
Settings → AI Connector → New connection.
| Field | Notes |
|---|---|
| Name | Name it after the machine or assistant you will use it from, so you know which to revoke later |
| Access | Standard, or Read only — read only can look at anything you can and change nothing |
The token is shown once. Copy it then; Sekee stores only a hash and a short prefix so you can identify the row afterwards.
The setup instructions appear when you click New connection — they are the same five steps whether or not you have just made a token, with the token filled in when you have.
Revoking
Revoke from the same screen. Every connection shows its prefix, when it was created, and when it was last used — "never used" is a useful thing to notice on a token that has been around for a month.
Revoking is immediate. You do not need to be an admin to manage your own connectors; they are your credentials, not organization content.
What an assistant can reach
Roughly: projects, directories, test cases and steps, shared steps, bug hunts and their cases and issues, sessions with their comments and bugs, requirements, automation runs and tests, errors, dashboards, and the rules below.
An area is only reachable once it has been deliberately added, its paths allow-listed, and its tenancy audited — so a newly built screen is not automatically exposed.
Rules an assistant reads first
See AI rules. They are standing instructions your organization sets, which the assistant is given before it writes anything.
Practical notes
- The transport is stateless streamable HTTP with JSON responses, not SSE.
- A POST needs
Accept: application/json, text/event-stream, or the SDK answers 406. This is the most common first failure when wiring a client up by hand. - To iterate on a client,
npx @modelcontextprotocol/inspectoris easier to debug against than a desktop assistant.