SurfacesMCP
MCP intelligence
Every MCP server your team's AI agents are routed through, with tool-level allow / warn / block counts and AI rationale per call.
You're viewing sample data. This is a live demo workspace pre-seeded with realistic incidents, tools, and policies.
Calls
2,067
Allowed
2,027
Warned
41
Blocked
40
Sensitive args
81
37 of 1,284 calls were blocked, all read_file attempts outside the project root. The policy is doing its job: agents read freely inside the repo and get stopped at the boundary. No path traversal got through.based on 2 sources
viaClaude CodeCursor·last seen 1m ago·8 users
1,284
calls
37
blocks
52
sensitive
642 calls, zero blocks. Reads and PR creation, all inside repos this org owns. Four calls touched a file with an embedded token, which the classifier flagged but the agent never exfiltrated. Healthy.based on GitHub MCP
viaClaude CodeCursor·last seen 5m ago·6 users
642
calls
0
blocks
4
sensitive
19 of 118 queries hit tables with customer data. Two were warned and one writer query was blocked. This server has direct database access and only an Unknown classification. I would set an explicit policy before usage grows.based on 2 sources
viaCursor·last seen 8m ago·2 users
118
calls
2
blocks
19
sensitive
New this window. One create_refund call was blocked. A coding agent reaching a live payments API is exactly the kind of access that needs a per-tool rule, not a server-wide allow. See the anomaly above.based on Stripe MCP
viaCursor·last seen 14m ago·1 user
23
calls
1
blocks
6
sensitive
© 2026 Northbeams IncStatus · Workspace dashboard · v1.0