> For the complete documentation index, see [llms.txt](https://help.tahua.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://help.tahua.io/product-news/2026/september/admin-read-mcp-self-service-connect.md).

# Self-service hosted MCP

## What it does

The hosted admin MCP lets an organisation connect an AI assistant (an MCP client such as Claude Code) to its read-only admin data. Previously, turning it on required an engineer in a Rails console: enable the tenant flag, create a public OAuth client, and register it. This feature makes the whole thing self-service for an integrations administrator.

## What's new

On **Settings → API Access → MCP connections**, an administrator with the *Manage Integrations* permission can now:

1. **Enable/disable hosted MCP access** for the tenant with a toggle (shown only where the platform has MCP enabled).
2. **Create a connection** — one click provisions the tenant's OAuth client and shows a **ready-to-paste connection config** (server URL + client id + callback port) to drop into their MCP client.
3. See the **authorized connections** and revoke any of them (as before).

To connect, the admin pastes the config into their client, runs the client's connect command (e.g. `/mcp` in Claude Code), and approves the connection in the browser. No console, no engineering ticket.

## How it works / notes

* **One client per tenant.** Provisioning is idempotent — clicking "Create connection" repeatedly reuses the tenant's existing client rather than creating duplicates. The OAuth application lives in the tenant's own schema.
* **The client id is not a secret.** The provisioned client is a public OAuth client whose redirect URI is restricted to loopback (`http://localhost`/`http://127.0.0.1`) on a fixed port, uses PKCE, and every connection still requires the administrator to sign in and consent in the browser. A leaked client id cannot be used to receive an authorization code on a remote server.
* **Disabling** the toggle immediately gates access: existing connections are rejected (and revoked) on their next request.
* **Loopback fix:** native OAuth clients that call back on `http://localhost:<port>` are now accepted in all environments (previously the platform forced HTTPS for the `localhost` host, which would have blocked MCP clients on staging/production).

## Security

Enabling MCP and minting a client are gated by the *Manage Integrations* permission, and (for provisioning) the per-tenant flag being on. Both actions are recorded as security events. The connection/consent path itself is unchanged — tokens remain bound to a single administrator, organisation, and resource.

## UI changes

The **MCP connections** settings page gains an enable/disable toggle, a "Create connection" button, and a connection-details panel with the copy-paste client config.
