tools/call — including shell commands — is classified and checked against your Sentrail workspace policies before it executes.
This is how rm -rf, git push --force, DROP TABLE, and other destructive commands get blocked or approval-gated. The cloud MCP gateway (mcp-gate) only sees JSON-RPC calls to a connected upstream server; it cannot intercept shell commands. The local interceptor can.
Quick start (under 2 minutes)
1. Install
2. Initialize
- Your Sentrail API key (
agk_…) - Your workspace ID
- The upstream MCP server URL the proxy should forward allowed calls to
3. Add the proxy to your agent
Paste the config snippet into your agent’s MCP settings:~/.claude/mcp_servers.json or .claude/mcp_servers.json in the project root.
For Cursor: Cursor Settings → MCP or ~/.cursor/mcp.json.
4. Start the interceptor
- Green → allowed, forwarded to upstream
- Red → blocked by policy
- Yellow → held for human approval
5. Verify with the built-in demo
rm -rf, git push --force, DROP TABLE, and cat README.md — and shows the Sentrail decision for each. No real commands are executed and no cloud connection is required.
How decisions are made
Supported agents
Any MCP-compatible client, including:- Claude Code
- Cursor
- Codex
- Custom agents using the MCP protocol over HTTP