Databricks
Databricks
Register twelfth’s MCP server in a Databricks
workspace as an MCP Service. Agents, AI Playground and other MCP clients in
Databricks can then read your twelfth workspace through Unity Catalog, with
Databricks governing who may call it.
The connection uses Dynamic Client Registration: Databricks registers
itself with twelfth automatically, and each Databricks user then signs in to
twelfth with their own account. Each connection reads as that person,
within their twelfth role and remits, and is read-only unless they turn on
Allow changes. No
twelfth key is created or copied.
This page is about Databricks reading twelfth. To have twelfth read data you
keep in Databricks, see the planned Databricks connector.
Before you start
Check these before anyone opens the registration form. The network policy step
is the one most often missed, and it needs a Databricks account admin.
In twelfth
- Everyone who will use the MCP Service needs an active twelfth account in a
twelfth workspace. They sign in with it once, from Databricks. If your
organisation has no workspace yet, book a demo.
In Databricks
- A workspace enabled for Unity Catalog, in a region that supports Model
Serving. - Users who register or sign in to the service need the Workspace access
entitlement. Users with only Consumer access can call a service shared
with them but can’t complete the per-user sign-in.
Unity Catalog privileges
| Who | What they do | Privileges |
|---|---|---|
| The admin who registers twelfth | Create the HTTP connection | CREATE CONNECTION on the schema that will hold it |
| The admin who registers twelfth | Create the MCP Service | USE CATALOG and USE SCHEMA on its catalog and schema, CREATE SERVICE on the schema, and USE CONNECTION on the connection |
| Everyone who uses twelfth from Databricks | Call the MCP Service | EXECUTE on the MCP Service, plus USE CATALOG and USE SCHEMA on its catalog and schema |
Grant users EXECUTE on the service, not USE CONNECTION on the connection.
Databricks warns that connection access lets a user call the server directly
and bypass the service’s tool selection, policies and auditing.
Allow api.twelfth.ai in the serverless network policy
Databricks calls external MCP servers from the workspace’s serverless compute.
If the workspace uses serverless egress control with Restricted access,
calls to twelfth are blocked until api.twelfth.ai is an allowed destination.
The registration form then shows:
PERMISSION_DENIED: Access to api.twelfth.ai is denied because of serverless network policy.
Nothing needs to change in twelfth. A Databricks account admin fixes it in
the account console (on Azure,
accounts.azuredatabricks.net):
- Find the policy attached to the workspace: open Workspaces, select the
workspace and read Network Policy. A workspace with no policy assigned
uses the account’s default policy. - Click Security, then the Networking tab. Under Policies, click
Context-based ingress & egress control and open that policy. - On the Egress tab, click Add destination above Allowed domains
and addapi.twelfth.ai. - Click Update.
Allowed-domain changes reach serverless compute within about ten minutes. Then
go back to the registration form and click Sign in to api.twelfth.ai again.
Prefer allowing the one domain. Switching the policy to Allow access to all
destinations (Full access) also works, but it opens the workspace’s
serverless compute to every site, and changing the access mode can take up to
24 hours to reach running workloads. A domain on the policy’s blocked list is
denied even under Full access.
If calls still fail, query the system.access.outbound_network system table
for denied destinations. Serverless egress control needs the Enterprise tier
(Premium on Azure), so workspaces on other tiers don’t see this error.
Databricks reference: MCP networking
and Manage network policies for serverless egress control.
Register twelfth as an MCP Service
Databricks moves these menus between releases. If a label differs, follow
Databricks’ current Register an external MCP server
guide; the values to enter stay the same.
- In the Databricks workspace, open Unity Gateway (AI Gateway in some
workspaces), select MCPs and add an MCP server (+ MCP →
Connect an existing MCP server, or Register MCP Server). You can also
go to Catalog, select a schema and click Create → MCP Service. - Choose the catalog and schema, and name the service, for example
twelfth.
The name can’t be changed later. - Click Create new connection and enter:
- Connection type: HTTP.
- Server URL:
https://api.twelfth.ai/mcp. If the form asks for the
parts separately, use Hosthttps://api.twelfth.ai, Port443
and Base path/mcp. - Auth type: Dynamic Client Registration. Leave OAuth scope
blank; twelfth’s discovery metadata supplies the rest.
- Under Tools, keep All tools or pick the ones to expose. A tool
twelfth doesn’t allow for the signed-in person is refused either way. - Click Sign in to api.twelfth.ai (or Login on the service’s page).
Sign in to twelfth, choose the workspace and approve the read tools. The
consent screen shows your Databricks workspace address as where the sign-in
returns; check it is yours. - Create the service. Its page lists twelfth’s tools once sign-in succeeds.
- On the service’s Permissions tab, grant
EXECUTEto the users, groups or
service principals who should use it. - Try it in AI Playground: add the service as a tool and ask for something
such as this week’s open actions.
Each colleague you grant access to signs in to twelfth once, with their own
account, the first time they call the service. Until they do, Databricks
returns an error asking them to authenticate.
Ending a connection
Each person’s sign-in is a separate connection in twelfth. End one from
Settings → AI & agents in twelfth; it stops on the next request. Revoking
EXECUTE or deleting the MCP Service in Databricks stops Databricks calling
twelfth for everyone.
Support
Email hello@twelfth.ai with the Databricks error
text, your twelfth workspace name and roughly when it happened.
Related
- Generic MCP: the endpoint, available tools and Allow changes
- AI & agents overview: every assistant and how it connects
- Databricks connector: the planned connector that reads your Databricks data into twelfth