@sentinel-atl/gateway

MCP Security Gateway — a drop-in proxy that adds identity, credentials, reputation, content safety, and audit to any MCP server.

Client → [Gateway: identity + credentials + reputation + safety + audit] → Server

Install

npm install @sentinel-atl/gateway

Quick Start

import { createGateway } from '@sentinel-atl/gateway';

const gw = await createGateway({
  name: 'my-gateway',
  enableSafety: true,
  minReputation: 30,
  rateLimitMax: 100,
});

// Block dangerous tools
gw.addToolPolicy('delete_all', { blocked: true });

// Require high reputation for admin tools
gw.addToolPolicy('exec_code', { minReputation: 80 });

// Process every incoming tool call
const result = await gw.processToolCall({
  toolName: 'search',
  callerDid: agent.did,
  credentials: [vc],
  authPayload: payload,
  authSignature: signature,
});

if (result.allowed) {
  const output = await upstreamServer.callTool('search', args);

  // Check output safety before returning
  const safety = await gw.checkResponseSafety(agent.did, 'search', JSON.stringify(output));
  if (safety.allowed) return output;
}

Features

| Feature | Description | |---------|-------------| | Per-caller rate limiting | Configurable max requests per window | | Tool policies | Block, scope-restrict, or reputation-gate any tool | | Pre-hooks | Custom async functions for per-tool authorization | | Content safety | Pre-dispatch and post-response content filtering | | Kill switch | Emergency revocation cascading to all delegates | | Reputation tracking | Automatic scoring with vouch support | | Offline mode | Degraded trust decisions when disconnected | | Full audit trail | Hash-chained, tamper-evident log of every call | | Stats & observability | Per-tool and per-caller metrics |

API

createGateway(config: GatewayConfig): Promise<MCPSecurityGateway>

Factory function that creates a fully initialized gateway.

gateway.processToolCall(request): Promise<GatewayResult>

Run the full Sentinel pipeline on an incoming tool call.

gateway.checkResponseSafety(callerDid, toolName, response): Promise<{allowed, safetyResult?}>

Check a tool response for content safety violations.

gateway.addToolPolicy(toolName, policy): void

Register per-tool policies (block, scopes, reputation, hooks).

gateway.revokeCaller(targetDid, reason): Promise<void>

Revoke a caller so they can no longer pass through the gateway.

gateway.killSwitch(targetDid, reason, downstreamDids?): Promise<KillSwitchEvent>

Emergency cascading revocation.

gateway.getStats(): GatewayStats

Per-tool and per-caller statistics.

License

Apache-2.0