Security
Tool safety
Skills and MCP connectors make AI useful at work, but each one is third-party code that can reach company data. Harriet reviews, contains, and audits every tool before it touches your systems.
The risk
Agentic AI has a supply chain now
Every skill is instructions your assistant may follow. Every connector is third-party code with access to documents, credentials, and internal systems. Approve them blindly and you have added a supply chain IT can't see.
Hardcoded bearer token in the MCP URL, and all tool calls routed through an external third-party service.
Trust
From "who installed this?" to "here's why we approved it"
Every extension gets a paper trail before it becomes part of how your company works. That means faster approvals for clean requests, and nothing to scramble for when compliance asks what changed.
FAQ
What security teams ask.
What does Harriet check before a skill or connector runs?
Every skill and MCP connector is threat-scored before it can reach company systems: risk signals like credential exposure, external routing, and unpinned packages surface up front, and nothing activates until a human reviews and approves it.
What happens when a scan flags something?
The tool is blocked pending a fix, with the findings listed against it, for example a hardcoded token or calls routed through an external third-party service. You can re-run a scan on demand after the issue is addressed.
How do we prove this to auditors?
Every approval, override, and tool call lands in an append-only audit trail, so when compliance asks what changed, the evidence already exists: who approved which tool, when, and why.
Bring the connector you would hesitate to approve
We'll show how Harriet turns it into a reviewable, contained, auditable decision.