Connect your AI agent over MCP.
Claude Desktop, Cursor or any MCP client can drive your InventDB workspace directly. It signs in as you, reads live data through typed tools, and every write it proposes waits for your approval.
What you are connecting.
Every workspace serves the Model Context Protocol on a single HTTPS endpoint. A connected client gets typed tools rather than a raw connection.
InventDB SOAR exposes 51 tools, including workflows and reports. InventDB Serverless exposes 40: the other 11 need AI inside the database, and InventDB Serverless has none. The agent can only call these tools, and every call runs as you, so your role, your row rules and your audit trail apply to it exactly as they apply to you.
Read
schema, describe_type, get, search and sql read structure and data.
Write
propose_insert, propose_update and propose_delete stage a change; commit applies it, cancel discards it.
Files
attach_search, attach_read, attach_download and upload work with attachments.
Reports
Generate and fetch reports from your own data.
The write tools are two-step by definition. A propose call returns a change for review; nothing touches the database until a commit. This is built into the tool definitions themselves, so it holds whether or not you configure anything.
Find your workspace URL.
Each workspace has its own address: the one in the browser bar when you sign in.
https://<your-workspace>.inventdb.comThe MCP endpoint is that address with /mcp on the end. Keep the full URL; you will paste it into your client in a moment.
If the address does not load: an idle workspace sleeps and wakes on the first request, so you may briefly see a waking screen. Let it finish before you connect a client, because a call made while the workspace is still starting can fail.
How sign-in works.
You never create or paste a token. The workspace implements OAuth 2.1 with PKCE and dynamic client registration, the flow MCP clients are built for.
The client registers itself, obtains its own access token and refreshes it when it expires. The first time you connect, it opens a sign-in page served by your own workspace; you log in there, review what the client is asking for, and approve it. The credential never enters the client's configuration file.
No stored password
Your password is typed into your workspace's sign-in page, never into a config file.
Tokens refresh themselves
Access tokens renew automatically, so nothing expires and quietly stops working weeks later.
Revoke one client at a time
The workspace exposes a standard revocation endpoint, so removing one client leaves your password and your other access untouched.
Scripts and CI are different. OAuth suits an interactive client that can open a browser. For unattended work, a scheduled job or a build step, use a service account instead: the credential then belongs to the job and can be revoked on its own rather than carrying a person's identity.
Point your client at it.
Give the client the endpoint and set its auth to OAuth. It discovers the sign-in and token endpoints from the workspace itself.
In the client's settings
In Claude Desktop this is Settings → Connectors → Add custom connector. Other clients have an equivalent remote MCP server screen. Enter the endpoint:
https://<your-workspace>.inventdb.com/mcpOr in a config file
Clients configured by file take the same two values:
{
"mcpServers": {
"inventdb": {
"url": "https://<your-workspace>.inventdb.com/mcp",
"auth": { "type": "oauth2" }
}
}
}Restart the client after editing its config file; none of them reload it live.
On the first connection a browser window opens on your workspace's sign-in page. Approve the request and the client stores its own token. There is no secret to copy into the file.
Verify the connection.
After you approve the sign-in, the client lists the tools it can see. That list tells you most of what you need to know.
51 or 40 tools listed
51 on InventDB SOAR, 40 on InventDB Serverless: the endpoint, the sign-in and your role all line up.
No browser window opened
The client is not treating this as an OAuth server. Check the URL ends in /mcp and that auth is set to OAuth rather than to a header or an API key.
Fewer tools than expected
Tools are filtered by role, so an account with narrower permissions sees a shorter list. That is how it is meant to work and not a fault.
Sign-in page did not load
An idle workspace may still be waking. Wait for it to finish, then reconnect.
Then ask the agent something only your data can answer. A record count per table is a good first question, because a wrong answer is obvious.
What the agent may and may not do.
The client signs in as you, so the agent operates under exactly your rules.
Your role
If you cannot delete records, neither can the agent.
Your row rules
Row-level security applies to every tool call, so the agent cannot read rows you cannot read, including through search and reports.
The audit trail
Every tool call is logged and attributed to you, like any other access.
Commits gate writes
A proposal is visible for review before anything is applied.
To give an agent less than you have, create a user with a narrower role and sign the client in as that user. The agent inherits the narrower view, and you can revoke that client without touching your own access.
That is the whole setup. The agent now works against the live database rather than an exported copy, and the same workspace keeps serving SQL over HTTP and REST at the same time.
See the full API surface