Integrate an agent or tool

Add Sernixa immediately before the code that performs a meaningful action. The SDK requests a decision, waits safely when review is required, and calls your function only after an approved outcome.

Install and configure

The TypeScript SDK supports Node.js 18 or later.

npm install @sernixa/sdk
export SERNIXA_BASE_URL=https://api.sernixa.com
export SERNIXA_API_KEY="<your-organization-api-key>"

Create the client from environment variables so credentials are not embedded in source code:

import { Client } from '@sernixa/sdk'

const client = new Client()

Before the first action, verify the credential and organization:

const identity = await client.whoami()
console.log(identity.organization_id, identity.permissions)

Gate a function

Wrap the smallest function that crosses a meaningful trust boundary. Use stable, policy-friendly context and keep unnecessary data out of metadata.

const governedRead = client.intercept(async (documentId: string) => documents.read(documentId), {
  intentId: 'customer-document-read',
  riskLevel: 'LOW',
  operationClass: 'read',
  dataSensitivity: 'confidential',
  systemsTouched: ['document-store'],
})

const document = await governedRead('doc_123')

The wrapped function remains yours. Validate its return value and handle its own business errors after Sernixa allows it to run.

Handle every outcome

Your integration should have an intentional path for:

Situation	Expected behavior
Approved	Execute once and validate the downstream result.
Pending review	Wait within the configured timeout without performing the action.
Rejected or blocked	Stop and show the reason to the operator.
Expired	Stop; submit a new request only when the action is still needed.
Authentication failure	Stop and repair credential delivery. Do not treat it as a policy result.
Network or service failure	Fail safely according to the action's risk; do not silently bypass governance.

Gate an MCP-style call

Ask Sernixa whether the proposed tool call may proceed, then let the MCP host dispatch an allowed call:

const decision = await client.mcpCallGate({
  mcpToolsetId: 'workspace-tools',
  toolName: 'read_file',
  intentId: 'read-workspace-report',
  arguments: { path: '/workspace/report.md' },
  riskLevel: 'LOW',
  operationClass: 'read',
  dataSensitivity: 'internal',
})

if (decision.status === 'accepted' || decision.status === 'auto_approved') {
  await mcpHost.callTool('read_file', { path: '/workspace/report.md' })
}

Sernixa governs the decision. The host owns the MCP connection, dispatch, timeout, response validation, and destination-system error handling.

Production checklist

Use a separate, least-privilege key for each application and environment.
Keep a stable agent identity, intent ID, and operation taxonomy.
Leave argument capture off unless review genuinely requires values.
Test approval, rejection, block, expiry, authentication failure, and network failure.
Make retries idempotent for mutable actions.
Verify a low-risk action in the Command Center before adding high-impact operations.
Record the downstream result when end-to-end completion matters.

Next step

Secure the integration with Authentication and API keys, then configure Policies and controls.