MCP Gateway

The Nexus MCP Gateway acts as a deterministic firewall between your probabilistic AI agents and your mission-critical infrastructure. It performs deep packet inspection on every JSON-RPC tool call before it reaches your backend.

Zero-LLM latency

Unlike prompt-based guardrails that rely on a secondary LLM to evaluate safety, the Nexus Gateway evaluates execution payloads against compiled Network Policy Language (NPL) rules entirely in-memory.

Metric	Value
p50 latency	~160 ms
p99 latency	< 201 ms
Token cost	0

Policies compile from the open npl-spec repository. View open-source spec →

Inspection pipeline

flowchart TD
  A[JSON-RPC request] --> B{Valid RPC 2.0?}
  B -->|no| X[BLOCK: MCP_MALFORMED_RPC]
  B -->|yes| C{Server whitelisted?}
  C -->|no| X2[BLOCK: MCP_SERVER_NOT_WHITELISTED]
  C -->|yes| D{Method blocked?}
  D -->|yes| X3[BLOCK: MCP_METHOD_BLOCKED]
  D -->|no| E{Destructive op?}
  E -->|yes| X4[BLOCK: MCP_DESTRUCTIVE_OP_BLOCKED]
  E -->|no| F{NPL policy}
  F -->|deny| X5[BLOCK + audit]
  F -->|require_human| H[Approval queue]
  F -->|allow| G[Forward to backend]

Decision types

Decision	Behavior
`allow`	Forward to backend (or mock in simulation mode)
`deny`	Block with stable `reasonCode`, audit before agent sees error
`require_human_approval`	Queue for operator review, block until resolved

Inspect an MCP call

curl -X POST "https://api.hardalion.com/api/v1/mcp/gateway/inspect" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "agt_prod_f8a92b",
    "execution_id": "exec_01hxyz",
    "server_id": "enterprise-db",
    "tool_name": "query_ledger",
    "execution_mode": "production",
    "request": {
      "jsonrpc": "2.0",
      "id": 1,
      "method": "tools/call",
      "params": {
        "name": "query_ledger",
        "arguments": { "account_id": "ACC-4421" }
      }
    }
  }'

Scope required: agents:execute.

Connecting a LangChain agent

Wrap your tools with the Nexus SDK and provide your AgentIAM passport:

import { NexusGateway } from '@hardalion/sdk'
import { myEnterpriseTools } from './tools'

const securedTools = NexusGateway.wrap(
  { invoke: myEnterpriseTools.invoke.bind(myEnterpriseTools) },
  {
    tenantId: 'ten_acme',
    agentId: 'agt_prod_f8a92b',
    apiKey: process.env.HARDALION_API_KEY,
    baseUrl: 'https://api.hardalion.com',
    policy: 'hardalion://nato-prou-strict-financial@1.0.0',
  },
)

// Pass securedTools to your LangChain or CrewAI instance.

Execution modes

Mode	Behavior
`production`	Inspect and forward allowed calls to real backends
`simulation`	Inspect and return mock responses (no backend contact)
`shadow`	Evaluate policies, log counterfactual blocks, do not enforce (see Shadow mode)

Policy pinning

Pin immutable policy versions with registry URIs:

hardalion://nato-prou-strict-financial@1.0.0

View bundled policies: NPL reference policies.

Gateway policies

Per-tenant MCP gateway policies configure:

Trusted server whitelist
Blocked methods and tools
Destructive operation denial
Data exfiltration heuristics

Configure via the Nexus console or tenant API.

OAuth 2.1 for remote MCP

Nexus is the authorization server for remote MCP servers. Agents do not carry static MCP keys. After inspect allows a call, the gateway mints a short-lived JWT whose aud is only that MCP server (mcp://{serverId}). Client-presented bearer tokens are rejected (MCP_TOKEN_PASSTHROUGH).

Discovery:

GET https://api.hardalion.com/.well-known/oauth-authorization-server

POST https://api.hardalion.com/oauth/register

Mint a token (client_credentials or authorization code after AgentIAM proof):

POST https://api.hardalion.com/oauth/token

Access tokens last at most 15 minutes. Set oauthMode: "mcp_as" on the gateway policy (or NEXUS_REQUIRE_MCP_OAUTH=true later) so production forwards include access_token bound to the target server.

Spawn a governed agent

Customers build the agent on Nexus. Voice, tools, AgentIAM, and MCP OAuth are issued together. A third-party "human agent" platform is not required for this path.

curl -X POST "https://api.hardalion.com/api/v1/agents/spawn" \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Sofia",
    "system_prompt": "Help the treasury desk reconcile cash.",
    "persona": { "tone": "warm", "language": "English" },
    "allowed_tools": ["query_ledger"],
    "mcp_audiences": ["mcp://core-banking"]
  }'

Scope required: agents:manage. The MCP client_secret is returned once.