Skip to content

Runtime authorization for AI agents

Vulnify allows, reviews, or blocks an agent action before it runs. The agent describes the action. Vulnify answers ALLOW, REVIEW, or BLOCK, and appends the decision to a hash-chained audit log.

Vulnify is a runtime authorization layer. It does not scan machines or source code for vulnerabilities. It governs what agents and automations can access and do. The web app is app.vulnify.io. The API is https://api.vulnify.io.

  1. 1

    Install the SDK

    Node.js: npm install @vulnify/sdk. Python: pip install vulnify. Create an API key in the app first. The full install is on the quickstart.

  2. 2

    Call check() or guard()

    check() returns the decision. guard() runs your function only when that decision is ALLOW. Run the export, send, or delete only after ALLOW.

  3. 3

    See the decision in the dashboard

    Open the event in app.vulnify.io. The overview counts live decisions. A TEST key stays in the sandbox and out of those charts.

const decision = await vulnify.check({
agent: 'SalesBot',
action: 'EXPORT_DATA',
resource: 'Customer Database',
destination: 'EXTERNAL_EMAIL',
recordsAffected: 12000,
});
if (decision.finalDecision === 'ALLOW') {
await exportCustomerRecords();
}

guard() is the shorter form. It throws for REVIEW and BLOCK. Both SDKs are on Node.js and Python.

Vulnify overview dashboard. The sidebar account email is hidden. Live decisions are counted here; this view has none in the selected period.

Node.js SDK

npm install @vulnify/sdk. check(), guard(), and the CLI. Node.js.

Python SDK

pip install vulnify. The same decision, with final_decision. Python.

Policies

Rules that allow, send to review, or block, in enforce or monitor mode. Policies.

Monitor mode

Record what would have been blocked or reviewed, and answer ALLOW. Monitor mode.

Audit log

A per-organization SHA-256 hash chain, and Verify audit chain in the app. Audit log.

Slack approvals

Approve or Deny a review from the alert. Slack.

API reference

POST /v1/events and the rest of the machine contract. API reference.

A check sends action metadata: the agent, the action, the resource, the destination, and how many records are affected. Vulnify is designed to store that metadata, not the contents of your records.

You can optionally send content so the server can scan it for sensitive data. Matches come back as findings. The content itself is not stored. See Risk engine.

Decision What you do
ALLOW Run the action.
REVIEW Do not run the action. A person has to approve it, for example in Slack.
BLOCK Do not run the action.

In monitor mode the event is stored and not enforced. Follow finalDecision. evaluatedDecision is what enforcement would have returned. If Vulnify cannot be reached, the SDKs fail closed by default and return BLOCK. A rejected request, including a body over 200 KB, throws instead. See Decisions.

Two ways to ask. Check, then act: your code calls the API or an SDK, and runs the tool only after ALLOW. Gateway: the app can register an HTTP API or an upstream MCP server as a route. Callers send the call to Vulnify. Vulnify decides, and forwards it to that upstream only when the decision allows it. The MCP gateway is that path for MCP tool calls. The Vulnify MCP server is a different host: tools for Cursor, Claude, and VS Code. The app describes the gateway path as beta. A credential for the upstream is injected only after ALLOW, and the agent never sees it.