🛡️ Secure. 📜 Auditable. 🙋 Human-Driven.

Permiscope is an open-source infrastructure layer that mediates all real-world actions performed by autonomous AI agents. Think of it as OAuth + Policy Engine + Audit System for AI agents.

🚀 Why Permiscope?

Without mediation, agents operate with full system permissions — a single bug or prompt injection can cause catastrophic damage. Permiscope enforces least privilege, human oversight, and auditability by default, making autonomous systems safe for production.

🧠 The Problem

The industry has moved beyond “Look what this agent can do” to “How do I stop this agent from breaking critical systems?”

Current agent frameworks typically operate with:

• ❌ All-or-nothing permissions
• ❌ Minimal oversight
• ❌ Limited traceability
• ❌ High operational risk

🛡️ The Solution: Mediated Agency

Instead of agents directly accessing files, APIs, or shells, every action flows through Permiscope’s Secure Execution Gateway.

graph LR
    Agent[🤖 AI Agent] --> Gateway[🔒 Secure Gateway]
    Gateway --> Policy{📜 Policy Engine}
    Policy -->|✅ Allow| System[💻 System / API]
    Policy -->|❌ Block| Agent
    Policy -->|🙋 Approvals| Human[👤 Human Admin]
    Human -->|Approve| System
    Gateway -.->|📝 Logs| Audit[Audit Trail]

✨ Key Capabilities

🚀 Quick Start

📦 Installation

npm install permiscope

⚡ 30-Second Quick Start

Get a guided tour of Permiscope in action:

npx permiscope --demo

This demo showcases allowed, blocked, and human-approved actions in a safe environment.

🔐 Security Configuration

Permiscope is designed for high-trust environments. Use these environment variables to harden your installation:

| Variable | Importance | Description | |----------|------------|-------------| | PERMISCOPE_AUDIT_SECRET | Critical | Secret key for HMAC-SHA256 audit log signing. | | PERMISCOPE_DASHBOARD_TOKEN | High | Bearer token required to update approvals via the API/Dashboard. | | PERMISCOPE_STRICT_LOGGING | Medium | Set to true to block actions if the audit log cannot be written. |

🛠️ Execution & CLI

Permiscope allows you to wrap agent commands safely:

• ✅ Allowed Action:

  permiscope run_command "echo hello"

• ❌ Blocked Action (Dangerous):

  permiscope run_command "rm -rf /"

💻 Web Dashboard

[!IMPORTANT] In production, always set PERMISCOPE_DASHBOARD_TOKEN for the control plane.

1. Start the Control Plane:

   # From within the repository
   npm run dev:dashboard

2. Access the Interface: Open http://localhost:3000 to review approvals and live audit trails.

🎯 Framework Agnostic by Design

Permiscope doesn't create agents — it governs what they can do.

Permiscope is a trust layer, not an agent framework. It wraps your existing execution logic with policy enforcement, approvals, and audit logging. Works with:

• LangChain / LangGraph
• CrewAI
• Custom agent loops
• Scripts and automations
• Any Node.js/TypeScript code

🏗️ Integration Guide

Primary API: PermiscopeAdapter

import { PermiscopeAdapter } from 'permiscope';

// Create a trust layer with default policy
const permiscope = new PermiscopeAdapter();

// Execute actions through the governed gateway
const content = await permiscope.act('read_file', { path: 'config.json' });

Wrapping Your Agent's Execution

Use wrap() to govern any function:

import { PermiscopeAdapter } from 'permiscope';
import * as fs from 'fs';

const permiscope = new PermiscopeAdapter({ policy: myPolicy });

// Wrap your existing executor
const safeReadFile = permiscope.wrap('read_file', async (params) => {
  return fs.readFileSync(params.path, 'utf-8');
});

// Now your function is governed
const content = await safeReadFile({ path: 'config.json' });

One-Liner Utility: withPermiscope

import { withPermiscope } from 'permiscope';

// Create a governed function in one line
const safeDelete = withPermiscope('delete_file', async (params) => {
  fs.unlinkSync(params.path);
}, { policy: strictPolicy });

await safeDelete({ path: '/tmp/temp.txt' });

Custom Policy Example

import { PermiscopeAdapter, defaultPolicy } from 'permiscope';

const permiscope = new PermiscopeAdapter({
  policy: {
    scopes: [
      ...defaultPolicy.scopes,
      { actionName: 'call_api', decision: 'REQUIRE_APPROVAL' }
    ]
  }
});

🔌 Integration Examples

With a Custom Agent Loop

import { PermiscopeAdapter } from 'permiscope';

const permiscope = new PermiscopeAdapter();

async function agentStep(action: string, params: Record<string, any>) {
  // All actions go through the trust layer
  return permiscope.act(action, params);
}

// Your agent loop
while (hasMoreWork) {
  const nextAction = await agent.plan();
  const result = await agentStep(nextAction.name, nextAction.params);
  await agent.observe(result);
}

With LangChain Tools (Conceptual)

import { PermiscopeAdapter, withPermiscope } from 'permiscope';

// Wrap LangChain tool executors
const safeShellTool = withPermiscope('run_command', async (params) => {
  return shellTool.call(params.command);
});

// Use in your chain
const tools = [safeShellTool, safeFileTool];

🔌 Framework Integrations

Permiscope works with any agent framework or custom workflow.

| Framework | Pattern | Demo | |-----------|---------|------| | LangChain | Wrap Tool Executors | langchain.ts | | CrewAI | Multi-Agent Governance | crewai.ts | | Custom Loops | Mediated act() calls | custom-loop.ts | | Workflows | Functional wrap() | workflow-runner.ts |

Learn more in the Integrations Guide.

🧪 Real-World Scenarios

Check out src/scenarios/ for full demos:

1. DevOps Agent: Safely edits configs, blocked from restarting services.
2. Data Agent: Reads raw data, writes processed output, blocked from overwriting raw.

📚 Documentation

See the /docs folder for detailed guides:

• Architecture — System design and trust boundaries
• Security Model — Threat model and guarantees
• Policy Engine — Rule evaluation and guardrails
• Approvals Workflow — Human-in-the-loop
• Custom Executors — Extending Permiscope
• FAQ — Common questions

🤝 Contributing

We welcome community contributions! Permiscope uses structured templates for Bug Reports, Feature Requests, and Pull Requests.

1. Fork the repository.
2. Follow the Contributing Guidelines.
3. Open a PR with our standardized template.

📄 License

MIT License. See LICENSE for more information.