MCP Tool Policies & Approvals
AVAILABLEEstablish enterprise governance, security rules, and human-in-the-loop approval workflows for tools and external Model Context Protocol (MCP) servers.
- Portal Location:
Projects > [Your Project] > Governance & Routing > MCP Tool Policies(/projects/[slug]/mcp-policies) - API Endpoint:
/v1/mcp/policies
Why MCP Tool Policies? #
Giving AI agents autonomous access to external systems (e.g., executing SQL updates, deleting files, sending emails, initiating payments) poses significant operational and security risks.
MCP Tool Policies enforce strict execution modes on a per-tool, per-role, or per-environment basis:
Rendering diagram...
graph TD
Agent["AI Agent invokes MCP Tool"] --> Gatekeeper["Naagmani Policy Gatekeeper"]
Gatekeeper --> Mode{"Tool Execution Mode?"}
Mode -->|"Auto-Approve (Read-Only)"| Run["Execute Immediately"]
Mode -->|"Require Approval (Mutating)"| Queue["Pause Execution & Alert Human Approver"]
Mode -->|Blocked| Reject["Reject Tool Call (HTTP 403)"]
Queue --> Decision{"Human Decision?"}
Decision -->|Approve| Run
Decision -->|Deny| Reject
Tool Execution Modes #
| Policy Mode | Behavior | Recommended Use Cases |
|---|---|---|
| Auto-Approve | The tool executes instantly without pausing the agent pipeline. | Read-only tools (get_customer, read_file, search_docs). |
| Require Approval | The request pauses and generates an approval ticket in the Developer Portal. The agent waits for a human admin decision. | Destructive or mutating tools (update_db, delete_repo, charge_card). |
| Disabled / Blocked | The tool cannot be invoked under any circumstances. | High-risk legacy or deprecated functions. |
Configuring Tool Policies in Developer Portal #
- In your project workspace, click Governance & Routing > MCP Tool Policies.
- Locate the target tool or MCP server connector.
- Set the default action (
Auto-Approve,Require Approval, orBlock). - Optionally configure Approval Expiration (e.g., reject tool call if no human responds within 15 minutes).