Run independently
No mandatory Shielda account or hosted service.
Inspect and adapt
Public source with Apache-2.0 or MIT licenses.
Choose your scope
Adopt the component that fits your application.
Run it. Inspect the result.
A boundary you can test.
Try two protection scenarios from the published Agent Guard SDK. Each tests a permitted action and an adversarial request against real library code.
- 01 / Scoped access
- A granted document read succeeds. An unauthorized resource is denied before its handler runs.
- 05 / Exact-action approval
- An approved action executes. A changed action or an expired approval is rejected.
Built-in fixtures use fake data, with no API key, model calls, or external side effects. They exercise the embedded library in one process; they do not establish host isolation or production protection.
1. Install in a new directory
npm install @shieldaai/agent-guard-scenarios@0.1.0-alpha.12. Save as check.mjs
import { runConformance } from '@shieldaai/agent-guard-scenarios';
const report = await runConformance({
scenarioIds: ['S01', 'S05'],
});
console.log(`${report.passedCases}/${report.caseCount} fixture checks passed`);
if (!report.passed) process.exitCode = 1;3. Run the checks
node check.mjsVerified against the published package ·
From the model call to the action.
Validate responses with SafeCall. Govern registered tool calls with Agent Guard. Track relevant dependency advisories with Security Watch. Each addresses a different part of the system.
Give agents scoped tool access, limit delegated permissions, and require approval for sensitive actions.
- Available today
- An embedded authorization core, an MCP tool adapter, and reproducible protection scenarios. Enforcement covers registered calls routed through the guard.
- In development
- A durable executor, isolated workloads, governed memory, and an agent access map are in development. The protected deployment is not yet released.
Apply Pydantic validation, consistent retries, known-pattern text checks, and call reports across your Python application.
- Available today
- A model-call wrapper for OpenAI, Anthropic, and custom providers. Tool permissions and process isolation remain the application's responsibility.
- In development
- Complete-response inspection before streaming output is released, plus usage reservations before calls, retries, and repairs. These additions are not yet published.
Bring Zod validation, consistent retries, known-pattern text checks, and call reports to your Node.js application.
- Available today
- A typed model-call wrapper for OpenAI, Anthropic, and custom providers. Text checks have defined coverage; they do not make every response safe.
- In development
- Buffered response validation and usage reservations, aligned with the Python SDK through shared behavioral fixtures. These additions are not yet published.
Security Watch
TypeScript · Track relevant vulnerabilities
Explore the repositoryMatch security advisories and news to your software inventory, keep the evidence, and send relevant findings to Slack or email.
- Available today
- Inventory-aware advisory monitoring, finding history, and queued notifications. Findings support human review; the tool does not automatically patch systems.
- In development
- Broader artifact and configuration tracking, bounded SBOM imports, and source coverage evidence are in development and not yet published.
Make agent authority visible.
Keep actions accountable.
We are developing stronger execution boundaries and a map of agent tasks, delegated access, and the resources they can reach. The aim is to help teams understand and govern agent activity across their organization.
The executor and inventory foundations have working implementations. Cross-platform release validation, the access-map interface, and broader organizational integration are still in development.
These open-source components sit alongside the Shielda platform, whose broader workflow connects environment context, reviewed action, and evidence. Platform connections are being developed separately; each public project remains independently usable.
Explore agent security