MCP servers,
behind the gateway.

Register CIRVIX as the gateway in .mcp.json and every JSON-RPC tool call is evaluated before the upstream server sees it. This is MCP security done at the enforcement point — the same AI agent security boundary, applied to every MCP server you connect. The full treatment is on the MCP security page.

Install

npm install -g @cirvix_ai/agent-control

Init

cirvix init

Connect

# register CIRVIX as the gateway for MCP servers in .mcp.json:
{ "mcpServers": { "cirvix": { "command": "cirvix", "args": ["gateway"] } } }

Policy

# declare which upstream tools are permitted; everything else is default-deny
cirvix policy check

Test — watch a decision

With Claude Code running under CIRVIX, ask it to read a credential file such as ~/.aws/credentials. The read is denied by policy before it happens, an audit event is written, and the agent sees a structured denial instead of your secrets.

Verify

Run cirvix logs to see the decision record: action normalized, rule matched, decision DENY, payload redacted. A legitimate workspace read shows PERMIT. If both behave that way, Claude Code is under control.

Install Free See the Attack Lab

MCP servers,behind the gateway.