- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
3 weeks ago
Hi Community,
I’m exploring how to connect Claude to a ServiceNow MCP server. In our setup, multiple employees may use the same Claude account and connect to the same ServiceNow instance.
I want to understand how to secure this setup and prevent misuse. For example:
How can I restrict MCP access so users can only perform approved actions, such as reading incident details?
How can I prevent unauthorized updates or deletions in ServiceNow?
Can I track which employee made each request if everyone uses the same Claude account and integration account?
Any recommended approach or implementation guidance would be appreciated.
Solved! Go to Solution.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
3 weeks ago
Hi Narayana,
The key thing to get right first: the identity boundary is ServiceNow, not Claude. Every MCP tool call arrives at the instance authenticated (OAuth) as a ServiceNow user, and the call executes in that user's context — roles, table/field ACLs, and audit (sys_audit, sys_journal, transaction logs) all apply to that authenticated user. Claude is just the client. That drives the answers to all three questions:
1. Restricting to approved actions (e.g. read incident details only)
- Only enable the tools you actually want in the MCP server configuration. If users only need to read incidents, expose the read/retrieval tools and do not enable create/update/delete-capable tools for that server/configuration.
- Treat tool selection as the outer gate and ACLs as the inner gate (see point 2). Both should agree.
2. Preventing unauthorized updates/deletes
- MCP does not bypass the platform security model. If the authenticated user only has read ACLs on incident (create/write/delete ACLs are separate grants — delete in particular is its own ACL operation), an update or delete attempted through MCP fails the same way it would via the UI or REST.
- So give MCP users a least-privilege role set: read on the tables they need, and no write/delete grants. Do not hand the MCP connection an admin or broad itil-style integration account "to make it work" — that is what turns a read use case into a misuse risk.
- If you have both read-only consumers and a genuine write use case, split them: separate MCP configurations/connections with different enabled tools and different user permissions, rather than one powerful shared setup.
3. Tracking which employee made each request
- With one shared integration account: you cannot. From ServiceNow's point of view there is exactly one user, so every audit record, Created by/Updated by and transaction log entry shows that account. Nothing in the Claude layer can be recovered inside ServiceNow afterwards.
- To get per-employee attribution, each employee must authenticate to ServiceNow as themselves (per-user OAuth from their own client configuration/login). Employees can still use a shared Claude subscription/seat arrangement on the client side if your licensing allows it — but the ServiceNow connection must be personal, so each tool call runs as that employee, is limited by their own ACLs, and is auditable to them.
- As a side benefit, per-user auth also fixes questions 1 and 2 per person: a read-only employee and a fulfiller with write access automatically get different effective MCP capabilities because their own permissions differ.
In short: shared ServiceNow integration account + shared identity = no attribution and a single point of excessive privilege. Per-user OAuth into ServiceNow, read-only tool selection for read use cases, and least-privilege ACLs is the pattern that answers all three of your questions.
Hope that helps with the design.
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
3 weeks ago
Hi Narayana,
The key thing to get right first: the identity boundary is ServiceNow, not Claude. Every MCP tool call arrives at the instance authenticated (OAuth) as a ServiceNow user, and the call executes in that user's context — roles, table/field ACLs, and audit (sys_audit, sys_journal, transaction logs) all apply to that authenticated user. Claude is just the client. That drives the answers to all three questions:
1. Restricting to approved actions (e.g. read incident details only)
- Only enable the tools you actually want in the MCP server configuration. If users only need to read incidents, expose the read/retrieval tools and do not enable create/update/delete-capable tools for that server/configuration.
- Treat tool selection as the outer gate and ACLs as the inner gate (see point 2). Both should agree.
2. Preventing unauthorized updates/deletes
- MCP does not bypass the platform security model. If the authenticated user only has read ACLs on incident (create/write/delete ACLs are separate grants — delete in particular is its own ACL operation), an update or delete attempted through MCP fails the same way it would via the UI or REST.
- So give MCP users a least-privilege role set: read on the tables they need, and no write/delete grants. Do not hand the MCP connection an admin or broad itil-style integration account "to make it work" — that is what turns a read use case into a misuse risk.
- If you have both read-only consumers and a genuine write use case, split them: separate MCP configurations/connections with different enabled tools and different user permissions, rather than one powerful shared setup.
3. Tracking which employee made each request
- With one shared integration account: you cannot. From ServiceNow's point of view there is exactly one user, so every audit record, Created by/Updated by and transaction log entry shows that account. Nothing in the Claude layer can be recovered inside ServiceNow afterwards.
- To get per-employee attribution, each employee must authenticate to ServiceNow as themselves (per-user OAuth from their own client configuration/login). Employees can still use a shared Claude subscription/seat arrangement on the client side if your licensing allows it — but the ServiceNow connection must be personal, so each tool call runs as that employee, is limited by their own ACLs, and is auditable to them.
- As a side benefit, per-user auth also fixes questions 1 and 2 per person: a read-only employee and a fulfiller with write access automatically get different effective MCP capabilities because their own permissions differ.
In short: shared ServiceNow integration account + shared identity = no attribution and a single point of excessive privilege. Per-user OAuth into ServiceNow, read-only tool selection for read use cases, and least-privilege ACLs is the pattern that answers all three of your questions.
Hope that helps with the design.