Connect Claude Code or Cursor to your database through MCP
TablePro
On this page
TablePro’s Model Context Protocol (MCP) server lets Claude Code and Cursor query your saved database connections without receiving their passwords. Enable the server, configure the client, then choose which connections it can read or write.
Configure the MCP client
#The bundled tablepro-mcp bridge starts TablePro if needed, finds its port and supplies the token. The client communicates with the bridge over stdin and stdout; its config needs only the executable path.
Claude Code takes it from the command line:
claude mcp add tablepro -- /Applications/TablePro.app/Contents/MacOS/tablepro-mcp
claude mcp list
The -- separates Claude Code's own flags from the command it runs.
Cursor reads ~/.cursor/mcp.json for every project, or .cursor/mcp.json in one project:
{
"mcpServers": {
"tablepro": {
"command": "/Applications/TablePro.app/Contents/MacOS/tablepro-mcp"
}
}
}
Restart Cursor after editing the file. If TablePro is installed somewhere other than /Applications, use that path.
Enable the server in TablePro
#- Open Settings > MCP and turn on Enable MCP Server. Status reads Running on port 23508, or the port it took instead when that one is busy.
- Click Connect a Client… and pick Claude Code or Cursor. The sheet shows the steps and a snippet with the path of your own install. TablePro does not edit other apps' config files, so paste it yourself.
- Ask the client to list TablePro's tools, or to call
list_connections. The list includeslist_connections,list_tables,describe_tableandexecute_query.
TablePro asks for approval on your Mac when a client first accesses a connection. Under Connection Access > Approval, choose Ask Once for Each Connection (default), Ask Every Time or Never Ask. This is a TablePro dialog, not a prompt the client can answer.
Set connection access
#Every saved connection has its own setting. Edit the connection, open the Options tab, and under Access set External Clients:
| Level | What a client can do |
|---|---|
| Blocked | Nothing. The connection is not even listed. |
| Read Only | Read the schema and run reads. Any write is refused before it reaches the database. This is the default. |
| Read & Write | Reads, and writes that pass Safe Mode. |
Leave production on Read Only or Blocked, and give the agent a read-only database role. Give Read & Write to a local or staging database where you want the agent to change data.
Require approval for writes
#A write from a client passes three checks, and the strictest one wins:
- External Clients on the connection, as above.
- The token's scope. The token
tablepro-mcpcarries is Read & Write. - Safe Mode, per statement, exactly as for a query you run yourself.
At each Safe Mode level, a write the first two checks allow does this:
| Level | An agent's write |
|---|---|
| Silent (the default) | Runs without asking. |
| Alert | Waits for you to confirm it in a dialog on your Mac that shows the whole statement. |
| Safe Mode | Waits for confirmation and Touch ID or your Mac password. |
| Read-Only | Is refused. |
Alert (Full) and Safe Mode (Full) also ask before reads. Set the level in the same Access section, or with the padlock in the toolbar while the connection is open.
With Read & Write and Silent, ordinary writes can run without a prompt. A DELETE without WHERE still raises the dangerous-query warning. Set Alert to review every write first.
DROP and TRUNCATE are handled apart from all of this. execute_query refuses them. They run only through the confirm_destructive_operation tool, which needs a Full Access token and your approval every time. The token tablepro-mcp carries is not Full Access, so an agent connected that way cannot drop or truncate anything.
Some statements are refused whatever the settings: anything that reads or writes files on the server or runs server-side code, and more than one statement in a single call, except a SQL Server script.
For a narrower token than the one tablepro-mcp carries, such as Read Only limited to a few connections, generate one under Settings > MCP > Authentication and connect the client over HTTP instead. The MCP Clients docs show that config.
Use a read-only database role
#Every check above runs in TablePro and judges a statement by its text. TablePro knows which built-in functions change data, such as nextval or set_config, but not what your own functions do: SELECT archive_orders() reads like a read. The database checks the role's privileges on every statement, whatever sent it.
For production, connect the agent as a role that can only read. On PostgreSQL, run the grants as the role that owns the tables, because ALTER DEFAULT PRIVILEGES covers only tables that role creates later:
CREATE ROLE agent LOGIN PASSWORD 'change-me';
GRANT CONNECT ON DATABASE app TO agent;
GRANT USAGE ON SCHEMA public TO agent;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO agent;
ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO agent;
On MySQL:
CREATE USER 'agent'@'%' IDENTIFIED BY 'change-me';
GRANT SELECT ON app.* TO 'agent'@'%';
Replace app with your database, and repeat the schema grants for each schema the agent should read. Then give the role its own connection:
- On the welcome window, right-click the production connection and choose Duplicate.
- Edit the copy to sign in as
agent, and set its External Clients to Read Only. - Set External Clients on the original to Blocked, so clients see only the copy.
To check the role, open the copy in TablePro and run a write that matches no rows:
UPDATE orders SET status = status WHERE false;
PostgreSQL answers permission denied for table orders, MySQL UPDATE command denied. If it runs, the role can write.
What a read-only role does not cover:
-
On PostgreSQL, a
SECURITY DEFINERfunction runs with its owner's privileges, and any role may call a function unlessEXECUTEis revoked fromPUBLIC. List them, and revokeEXECUTEfromPUBLICon any that write:SELECT p.oid::regprocedure FROM pg_proc p JOIN pg_namespace n ON n.oid = p.pronamespace WHERE p.prosecdef AND n.nspname NOT IN ('pg_catalog', 'information_schema'); -
On MySQL, grant nothing beyond
SELECT. A routine runs as its definer by default, soEXECUTEwould let the agent call one that writes. -
ALTER ROLE agent SET default_transaction_read_only = onis not a boundary: the session canSETit back.
Review activity
#View Activity… in Settings > MCP lists every tool call and query, with the connection and the outcome. Statements there are stored as a SHA-256 digest, not as text. Queries a client runs also appear in the query history of that connection, unless you turn off Log MCP queries in history.
Limitations
#- The server listens on
127.0.0.1only. There is no remote mode; a client on another machine needs an SSH port forward you set up and own. - It runs in the Mac app. The iPhone and iPad app has no MCP server.
- Clients see your saved connections, never their passwords, and cannot create or edit connections or change Safe Mode.
- A result stops at Default row limit, 500 rows unless you change it, and a client can ask for more only up to Maximum row limit, 10,000 by default.
- A call waits at most Query timeout, 30 seconds unless you change it, and a client can ask for up to 300. A read still running then is cancelled. All three limits are in Settings > MCP.
Troubleshooting
#- TablePro is not running: the bridge could not start the app or find the server, and gave up after 10 seconds. Open TablePro and check that Status says running.
- This connection is read only for external clients: the statement writes and External Clients is Read Only. Change it, or run the statement in TablePro yourself.
- The connection is missing from
list_connections: it is Blocked, or its AI policy is set to Never.
See MCP Tools for the tool list and MCP Server for settings. Feature overviews: MCP and Safe Mode. If TablePro cannot reach the database, check SSH tunneling or Docker port publishing.
Related pages
Available for Mac, Linux, iPhone and iPad.
Download