@opena2a/aim-sdk

AIM SDK for TypeScript/Node.js

Official TypeScript SDK for Agent Identity Management (AIM) - secure identity verification for AI agents.

Managed hosting available at aim.opena2a.org/get-started. Self-host instructions below.

Installation

npm install @opena2a/aim-sdk
# or
yarn add @opena2a/aim-sdk
# or
pnpm add @opena2a/aim-sdk

Quick Start

import { AIMClient, AgentType } from '@opena2a/aim-sdk';

// Create client
const client = new AIMClient({
  baseUrl: 'https://aim.example.com',
  apiKey: process.env.AIM_API_KEY,
});

// Register an agent
const agent = await client.registerAgent({
  name: 'my-ai-agent',
  displayName: 'My AI Agent',
  agentType: AgentType.LANGCHAIN,
  capabilities: ['file:read', 'api:call'],
});

// Verify an action before execution
const result = await client.verifyAction({
  action: 'file:read',
  resource: '/data/config.json',
});

console.log(`Action allowed: ${result.actionAllowed}`);
console.log(`Trust score: ${result.trustScore}`);

Features

• Full TypeScript Support: Complete type definitions for all APIs
• Ed25519 Signatures: Cryptographic signing for secure verification
• OAuth 2.0: Automatic token management with client credentials flow
• Express Middleware: Easy integration with Express.js applications
• Fastify Plugin: First-class support for Fastify applications
• Automatic Retries: Built-in retry logic with exponential backoff
• Local Credential Verification: Verify signed ATX credentials offline against cached trust anchors, no per-action call to a central service
• Delegation Chains: Create and verify Ed25519 delegation chains (cross-engine interop), with signature, identity, scope-narrowing, and expiry all enforced at verification time

Express Integration

import express from 'express';
import { createAIMMiddleware, verifyAction, aimErrorHandler } from '@opena2a/aim-sdk/express';

const app = express();

// Add AIM middleware globally
app.use(createAIMMiddleware({
  baseUrl: 'https://aim.example.com',
  apiKey: process.env.AIM_API_KEY,
}));

// Verify specific actions on routes
app.post('/api/data',
  verifyAction('data:write'),
  (req, res) => {
    // Action has been verified
    res.json({ success: true });
  }
);

// Access AIM context in handlers
app.get('/api/profile', (req, res) => {
  const { agentId, trustScore } = req.aim ?? {};
  res.json({ agentId, trustScore });
});

// Optional: map SDK errors thrown in later handlers to HTTP responses
// (ActionDeniedError -> 403, AuthenticationError -> 401). aimErrorHandler IS
// the four-argument handler — pass it to app.use, do not call it.
app.use(aimErrorHandler);

Fastify Integration

import Fastify from 'fastify';
import { aimPlugin, verifyAction } from '@opena2a/aim-sdk/fastify';

const fastify = Fastify();

// Register the AIM plugin
await fastify.register(aimPlugin, {
  baseUrl: 'https://aim.example.com',
  apiKey: process.env.AIM_API_KEY,
});

// Verify actions with preHandler hook
fastify.post('/api/data', {
  preHandler: verifyAction('data:write'),
}, async (request, reply) => {
  return { success: true };
});

Local Credential Verification (offline)

An ATX (Agent Trust eXtension) credential is a signed, portable credential designed to be verified locally: the signature is checked against the issuer's cached public key in roughly a millisecond, the issuing node is never on the verification path, and revocation rides on an asynchronously-refreshed, short-lived cached list. When you configure localVerification with cached trust anchors, the client verifies a resolved credential offline and decides authorization from the credential's own signed claims — no per-action call to a central service. The remote verifyAction POST is retained as the fallback for when no local credential is available.

The verifier is the shared, conformance-locked @opena2a/atx-verify — byte-for-byte interoperable with the Go (opena2a-registry/pkg/atcverify) and Python reference verifiers.

import { AIMClient } from '@opena2a/aim-sdk';

const client = new AIMClient({
  // Cached once from AIM/the Registry; refresh the CRL off the hot path.
  localVerification: {
    trustedIssuers: ['did:opena2a:issuer-1'],
    publicKeys: [{ algorithm: 'Ed25519', publicKeyHex: '<issuer raw ed25519 pubkey hex>' }],
    // crl: { entries: [{ agentId, reason }] }, // optional cached revocation list
  },
});

// Resolve the agent's ATX once (the AAP broker / network step), then cache it.
// `resolvedAtx` is the signed ATX credential returned by that broker/network step.
const resolvedAtx = await fetchAtxForAgent(); // your resolution step
client.setLocalCredential(resolvedAtx);

// Per action: verified offline, sub-millisecond, no network.
const result = await client.verifyActionLocally({ action: 'file:read' });
// or just call verifyAction(): it takes the local path automatically when a
// credential is cached, and falls back to the remote POST otherwise.

Authorization is gated on signed capabilities. ATX v1.1 credentials carry capabilities under the signature, so they are trusted; v1.0 capabilities are forgeable by the holder and are refused by default. Passing requireSignedCapabilities: false to LocalVerifier.authorize overrides this, but then authorization runs on holder-forgeable capabilities — only do this for a closed, trusted v1.0 deployment, never across a trust boundary.

Multi-issuer anchor sets. Give each key a DID-URL keyId (e.g. did:opena2a:authority:opena2a.org#key-1). @opena2a/atx-verify binds a key to its controller DID, so a key may only verify credentials issued by that DID (or, for v1.1, a signed issuerChain authority) — one trusted issuer cannot impersonate another. A key with no keyId fragment is unbound and eligible for any issuer: fine for a single-issuer anchor set, unsafe for a multi-issuer one.

For credential verification without the action-authorization adaptation, use the verifier directly:

const verifier = client.getLocalVerifier();
const { valid, context, rejectCategory } = await verifier!.verifyCredential(atx);

Network is reserved for credential resolution (the AAP broker hands the agent its ATX) and the periodic CRL refresh — never for a per-action decision.

Delegation Chains

A delegation is a signed statement that one key (the delegator) grants a set of scopes to another key (the delegate), optionally chained so authority passes from a root through intermediaries to a leaf. Delegations use did:key (Ed25519) and a canonical signing form for cross-engine interop.

import {
  generateKeyPair,
  createDelegation,
  verifyDelegation,
  verifyDelegationChain,
} from '@opena2a/aim-sdk';

const root = await generateKeyPair();
const coordinator = await generateKeyPair();
const worker = await generateKeyPair();

// Root delegates to a coordinator, which sub-delegates a narrower scope set.
const d1 = await createDelegation({
  delegatorKeyPair: root,
  delegatePublicKey: coordinator.publicKey,
  scopes: ['search', 'memory.read', 'memory.write'],
  // expiresAt defaults to 7 days from now; pass an ISO-8601 string to set it.
});

const d2 = await createDelegation({
  delegatorKeyPair: coordinator,
  delegatePublicKey: worker.publicKey,
  scopes: ['search', 'memory.read'], // must be a subset of the parent's scopes
  parentDelegation: 'd1',
  parentExpiresAt: d1.expiresAt,     // a child must not outlive its parent
});

const { valid, results } = await verifyDelegationChain([d1, d2]);
// valid === true only if every hop passes: signature, delegator identity,
// scope narrowing, chain linkage, trust attenuation, AND temporal validity.

verifyDelegation and verifyDelegationChain enforce the delegation's signed createdAt/expiresAt window. An expired delegation is rejected, and verification fails closed on a missing, unparseable, or inverted (createdAt after expiresAt) timestamp. A chain is evaluated against a single instant so every hop is judged by the same clock, and a child that outlives its parent is rejected even while the parent is still live (a delegate cannot hold authority in time beyond its delegator). When creating a sub-delegation, pass parentExpiresAt so the child's default expiry is capped at the parent's and an over-long child expiry is refused at creation.

Pass an explicit evaluation time for deterministic tests or offline / as-of verification:

// Verify as the chain would have stood at a specific instant.
await verifyDelegation(d1, { verifyAt: '2026-07-15T00:00:00.000Z' });
await verifyDelegationChain([d1, d2], { verifyAt: new Date('2026-07-15T00:00:00Z') });

For the raw signature check without temporal evaluation (archival or audit inspection where authenticity matters independent of time), use verifyDelegationSignature; for the standalone temporal check and its reason string, use checkDelegationTemporalValidity:

import { verifyDelegationSignature, checkDelegationTemporalValidity } from '@opena2a/aim-sdk';

await verifyDelegationSignature(d1);                       // true if the signature is authentic
checkDelegationTemporalValidity(d1);                       // { valid, error? } at the current time
checkDelegationTemporalValidity(d1, '2026-07-15T00:00:00.000Z');

Causal-Denial Telemetry (opt-in)

The SDK can correlate why a blocked agent action happened by joining three signals around one verified action: the authorization outcome (an observed fact), the classified intent, and the injection cause (both inferences). The full correlated record is authoritative and stays on the machine; only an anonymized shared indicator is ever uploaded. This section covers the causal-denial channel only — the runtime-protection module ships a separate structural-signature channel that is on by default; see the Runtime Protection section for its scope and opt-out.

The causal-denial channel is off by default and gated by two independent opt-ins:

• Capture (telemetry.enabled) — mints a correlation ID per verifyAction, assembles records, and appends them to a local log at ~/.opena2a/correlated-events.jsonl. Nothing leaves the machine.

• Share (telemetry.relay.enabled) — a best-effort relay reduces local records to anonymized indicators and uploads only denied_injection_attempt events to the public, count-only Registry endpoint (POST /api/v1/telemetry/runtime).

  The shared indicator carries no payloads, paths, credentials, resource/capability names, correlation ID, agent ID, or denial reason text. It carries only: a validated Threat Matrix techniqueId (T-NNNN, dropped if malformed) and its source, a detection confidence, the coarse enforcement outcome (deny/allow), the integrator's self-declared packageName/ agentCategory, daySinceInstall, runtimeEnv, triggeredAt, and a sensorToken — a stable per-device pseudonym (sha256(host+user+local salt)) that lets the Registry de-duplicate without identifying you.

const client = new AIMClient({
  baseUrl: 'https://aim.example.com',
  apiKey: process.env.AIM_API_KEY,
  telemetry: {
    enabled: true,                 // stage 1: capture records locally
    relay: {
      enabled: true,               // stage 2: upload anonymized indicators
      packageName: 'acme/support-agent', // your app's public sensor label
      packageVersion: '2.1.0',
      // registryUrl defaults to https://api.oa2a.org
      // intervalMs defaults to 60000, batchSize to 50
    },
  },
});

// Per-action inputs (populated by the runtime-protection module, or supplied
// directly). Ignored unless telemetry is enabled.
await client.verifyAction({
  action: 'file:read',
  resource: '/data/sensitive.json',
  telemetry: {
    intent: { intentClass: 'exfiltration', confidence: 0.7, blocked: true, source: 'nanomind-intent' },
    detection: {
      injectionDetected: true,
      techniqueId: 'T-2002',
      techniqueSource: 'interim-mapping',
      confidence: 0.84,
      detector: 'nanomind-guard',
      detectedAt: new Date().toISOString(),
    },
  },
});

// Stop the internally managed flush timers when shutting down.
client.closeTelemetry();

Guarantees. Telemetry runs off the enforcement path and is best-effort: a capture, join, or upload failure is swallowed and never changes an action's verification result. When both opt-ins are off, no correlation ID is minted, no header is attached, and nothing is written or sent.

TelemetryConfig fields: enabled (capture switch), enforcementSource (stamped on the enforcement fact, default aim-pdp), joiner (supply your own to control the sink/lifecycle), and relay (RelayConfig: enabled, registryUrl, packageName, packageVersion, agentCategory, dataDir, batchSize, intervalMs, timeoutMs). The CorrelatedRelay class is also exported for standalone use (e.g. draining the local log from a CLI).

Runtime Protection (@opena2a/aim-sdk/arp)

The SDK ships the ARP (Agent Runtime Protection) engine as a subpath module. It observes an agent process from the inside — monitors and interceptors for process, network, filesystem, prompt, MCP, and A2A activity; a rule-based event engine (L0); a behavioral anomaly twin (L1); and an intelligence coordinator (L2) — and produces the detection inputs shown in the telemetry section above.

The boundary with the hackmyagent scanner: scan-time analysis (static scanning, hardening rules, artifact classification) lives in hackmyagent; runtime protection lives here, inside the agent process. This module is the canonical home of the ARP engine; hackmyagent's ./arp export is being converted to a thin re-export of it.

import { EventEngine, FilesystemMonitor, EnforcementEngine } from '@opena2a/aim-sdk/arp';

const engine = new EventEngine(config);
const monitor = new FilesystemMonitor(engine);
await monitor.start();

Interception scope. The filesystem and process interceptors patch the CJS module registry (require('fs'), require('child_process')). Code that loads those builtins via require or an ESM namespace/default import (import fs from 'fs') is observed. Code that captured named ESM bindings before the interceptor started (import { readFileSync } from 'fs') bypasses them — ESM bindings are resolved at link time and cannot be patched. The monitors (which poll rather than intercept) are unaffected. Start interceptors as early as possible in the process, and treat them as one observation layer, not a sandbox.

Classification of runtime events is supplied through the injectable ClassificationProvider seam. The default NanoMindGuardClassificationProvider talks to the local NanoMind-Guard daemon over a Unix socket and only accepts signed classification results; when the daemon is absent the annotator degrades to "no classification" — it never fabricates a label and never blocks.

Invariant: the runtime-protection module is a telemetry producer. Its detection outputs flow through the telemetry.detection seam into the correlated record; they never enter verifyAction's allow/deny decision.

Structural signature telemetry (on by default, opt-out). Unlike the opt-in causal-denial channel above, the ARP engine shares structural attack signatures by default: when a detection fires, the structural shape of the event sequence — technique identifier, event-type sequence, severity, and a one-way hash of structural tokens — is signed and reported to the OpenA2A registry (https://api.oa2a.org), so an attack shape first seen at one deployment can protect others. Prompts, model responses, tool arguments, file contents or paths, command lines, environment values, secrets, IP addresses, hostnames, and account, tool, agent, and model names are never shared. A one-time disclosure is printed before first collection, and every payload is appended to a local audit log (~/.opena2a/telemetry-audit.log, JSONL) before it is sent. Opt out with any one of:

• OPENA2A_TELEMETRY_OPTOUT=1 (or ARP_TELEMETRY_DISABLED=1) in the environment,
• signatureTelemetry: { enabled: false } in your ARP config, or
• writeOptOutMarker() from @opena2a/aim-sdk/arp, which persists ~/.opena2a/telemetry-optout across processes.

Any one of these is the runtime-protection module's master switch: it disables every telemetry channel the module can produce (structural signatures and the opt-in legacy GTIN runtime channel). The causal-denial channel above is controlled solely by its own telemetry client config and is off unless you enabled it. To also delete signatures this sensor already shared, call purgeRemoteSignatures() (right-to-delete; best-effort, never blocks the local opt-out).

Configuration

Environment Variables

The SDK automatically reads from these environment variables:

Variable	Description
AIM_BASE_URL	AIM server base URL
AIM_API_KEY	API key for authentication
AIM_ORGANIZATION_ID	Organization ID
AIM_AGENT_ID	Pre-registered agent ID
AIM_PRIVATE_KEY	Ed25519 private key (base64)
AIM_PUBLIC_KEY	Ed25519 public key (base64)
AIM_DEBUG	Enable debug logging (true/false)

Client Options

const client = new AIMClient({
  baseUrl: 'https://aim.example.com',  // AIM server URL
  apiKey: 'your-api-key',              // API key
  organizationId: 'org-uuid',          // Organization ID
  autoRegister: true,                  // Auto-register if not registered
  timeout: 30000,                      // Request timeout in ms
  debug: false,                        // Debug logging
  headers: {},                         // Custom headers
});

Agent Types

The SDK supports various agent types:

import { AgentType } from '@opena2a/aim-sdk';

// LLM Providers
AgentType.CLAUDE
AgentType.GPT
AgentType.GEMINI

// Frameworks
AgentType.LANGCHAIN
AgentType.CREWAI
AgentType.AUTOGEN

// Assistants
AgentType.COPILOT
AgentType.ASSISTANT

Error Handling

import {
  ActionDeniedError,
  AuthenticationError,
  RateLimitError,
} from '@opena2a/aim-sdk';

try {
  await client.verifyAction({ action: 'file:delete' });
} catch (error) {
  if (error instanceof ActionDeniedError) {
    console.log(`Denied: ${error.reason}`);
    console.log(`Trust score: ${error.trustScore}`);
  } else if (error instanceof RateLimitError) {
    console.log(`Retry after: ${error.retryAfter} seconds`);
  }
}

Class identity is per entry point. Each entry point (., /arp, /express, /fastify) bundles its own copy of the error classes, so instanceof only matches errors raised by the same entry point you imported from. When you use an integration, import the error classes from that same integration (@opena2a/aim-sdk/express and /fastify re-export the full error family); for checks that must work across entry points, match on error.code (e.g. 'ACTION_DENIED') instead.

Credential Management

import {
  loadCredentialsFromFile,
  saveCredentialsToFile,
} from '@opena2a/aim-sdk';

// Save credentials after registration
const credentials = client.getCredentials();
await saveCredentialsToFile(credentials, '.aim/credentials.json');

// Load credentials on startup
const saved = await loadCredentialsFromFile('.aim/credentials.json');
client.setCredentials(saved);

API Reference

AIMClient

• registerAgent(options) - Register a new agent
• verifyAction(options, atx?) - Verify an action (local path when a credential is cached, remote POST fallback otherwise)
• verifyActionLocally(options, atx?) - Verify an action against a locally-held ATX credential, fully offline
• setLocalCredential(atx) - Cache the resolved ATX credential for offline verification (pass null to clear)
• getLocalVerifier() - The configured LocalVerifier, or null when local verification is not enabled
• getAgent() - Get current agent info
• updateAgent(updates) - Update agent metadata
• reportCapabilities(capabilities) - Report agent capabilities
• getTrustScore() - Get current trust score
• getCredentials() - Get stored credentials
• setCredentials(credentials) - Set credentials

Types

See src/types/index.ts for complete type definitions (source is not shipped in the npm package; the bundled .d.ts files carry the same types).

License

Apache-2.0

Contributing

See CONTRIBUTING.md for guidelines.