Intent-drift firewall for autonomous AI agents. Every tool call is evaluated against a baseline; decisions (allow / warn / block) come with full explanations. Built for trust, adoption, and audit.

1. Agents act — Your agent calls tools. NeoLayer does not decide which tools.
2. NeoLayer evaluates — Each call is evaluated for intent drift against a baseline.
3. Policy intervenes — Allow, warn, or block based on drift and sensitivity.

See WHY_NEOLAYER.md for risk mitigation and design principles.

npm install
cp .env.example .env   # set GEMINI_API_KEY from https://aistudio.google.com/apikey
npm run verify-env && npm run verify-db
npm start
# Open http://localhost:3000 → Run Simulation

See QUICKSTART.md for detailed setup.

Integrate in under 30 minutes. See docs/integrations/.

Hello World (Node.js):

import { IntentGuardWrapper } from './src/wrapper/intentguard-wrapper.js';

const wrapper = new IntentGuardWrapper({
  apiUrl: 'http://localhost:3000',
  agentName: 'my-agent',
  baselineIntent: 'Help users with support tickets'
});
await wrapper.initialize();

const safeSearch = wrapper.wrapTool('search_tickets', searchTickets, {
  sensitivity: 'medium',
  declaredPurpose: 'Find open tickets'
});

const tickets = await safeSearch({ status: 'open' });

What gets intercepted: Tool name, params, sensitivity, declared purpose.
On block: Tool does not run; IntentGuardBlockError thrown.
On warn: Tool runs; warning logged; human review recommended.

• Wrapper SDK — Intercept tool calls; enforce allow/warn/block.
• Intent inference — Gemini-based with semantic context; no silent fallbacks.
• Policy + explainability — Risk breakdown, human-readable explanation, contributing signals.
• Dry-run — POST /api/policy/dry-run to see what NeoLayer would do without enforcement.
• Session risk summary — GET /api/sessions/:id/risk-summary for executive view.
• REST API — Sessions, tool-calls, intent, policy; simulation and trace replay.
• Docker — Dockerfile and docker-compose.

• Node.js ≥ 18
• SQLite (better-sqlite3)
• Gemini API key (required for intent inference)

See QUICKSTART.md for full API list.

• WHY_NEOLAYER.md — Problem, core ideas, what NeoLayer mitigates
• QUICKSTART.md — Setup and first run
• ARCHITECTURE.md — Simplified architecture
• DEBUGGING.md — Gemini failures, policy tuning, logs
• docs/integrations/ — Node, Python, LangGraph integration guides

MIT.