Skip to main content
Sentrail offers two connection modes for Claude Code. Choose Option A if you want to intercept local shell commands (rm -rf, git push --force, DROP TABLE, etc.). Use Option B if you only need to govern MCP tool calls routed to an upstream server.
For rm -rf, git push --force, DROP TABLE, and other shell-level commands, you must use Option A (local interceptor). The cloud proxy only sees MCP JSON-RPC calls — it cannot intercept commands the agent runs via bash or a terminal tool.
The Sentrail CLI runs a local proxy at http://localhost:3773. Claude Code connects to it instead of the upstream MCP server. Every tools/call — including shell commands — is classified and checked against your policies before execution.

​
1. Install the CLI

npm install -g sentrail

​
2. Initialize

sentrail init
Enter your API key (agk_…), workspace ID, and upstream MCP server URL when prompted. The command writes ~/.sentrail/config.json and prints the MCP snippet for the next step.

​
3. Add to Claude Code config

Add the following to your Claude Code MCP configuration. The location depends on your setup:
  • Global: ~/.claude/mcp_servers.json
  • Project: .claude/mcp_servers.json in the project root
{
  "mcpServers": {
    "sentrail": {
      "url": "http://localhost:3773"
    }
  }
}

​
4. Start the interceptor

sentrail start
The proxy runs in the foreground and logs every decision with color coding (green = allow, red = block, yellow = pending approval).

​
5. Verify with the demo

sentrail demo
The demo simulates four agent actions — rm -rf, git push --force, DROP TABLE, and cat README.md — and shows the decision for each. No real commands are executed.
Do not commit your API key to version control. The ~/.sentrail/config.json file is local to your machine.

​
Option B — Cloud MCP proxy

Point Claude Code directly at the Sentrail mcp-gate cloud endpoint. This governs MCP tools/call requests to an upstream MCP server but does not intercept local shell commands.

​
Prerequisites

  • A Sentrail workspace with an MCP server connected (Tools → MCP → Connect)
  • An API key with the mcp scope

​
Configuration

{
  "mcpServers": {
    "sentrail": {
      "url": "https://<project-ref>.supabase.co/functions/v1/mcp-gate/sse",
      "headers": {
        "Authorization": "Bearer agk_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
        "X-Agent-Id": "claude-code"
      },
      "transport": "sse"
    }
  }
}

​
Verifying the connection

After saving the config, restart Claude Code and run: 
/mcp
You should see sentrail listed as a connected server with the tools from your upstream MCP server.

​
How it works

Claude Code sends all MCP calls to mcp-gate instead of your upstream server:
  1. initialize — mcp-gate responds with Sentrail server info and protocol version 2024-11-05
  2. tools/list — mcp-gate fetches the tool list from your upstream server (cached 60 seconds) and returns it unchanged
  3. tools/call — mcp-gate evaluates the call against your policies, then either forwards it, blocks it, or defers it
When a call requires approval, Claude Code receives a result with _meta.sentrail.status = "require_approval" and a message explaining that approval is needed. Claude Code will surface this to you. Poll the status endpoint or wait for the dashboard notification.

​
Agent identity

The X-Agent-Id: claude-code header identifies this connection in audit logs. You can create agent-specific policies using this ID: 
{
  "name": "Allow Claude Code to create issues",
  "tool": "mcp",
  "actionPattern": "*",
  "toolNamePattern": "create_issue",
  "conditions": { "allowedAgents": ["claude-code"] },
  "mode": "allow",
  "priority": 10
}

​
Troubleshooting

Check that the Sentrail MCP endpoint ends with /mcp-gate/sse. stdio is not supported by mcp-gate.
Verify the API key is valid, not expired, and has the mcp scope. Check Settings → API Keys in the dashboard.
Go to Tools → MCP in the dashboard and connect an upstream MCP server before using mcp-gate.