🔐 VaultAgent

Permission control for AI Agents — control what your agents can and cannot do.

VaultAgent is an open-source SDK + cloud dashboard that gives you fine-grained permission control over AI agent tool calls. Think of it as IAM for AI Agents.

The Problem

AI Agents can call tools — read files, send emails, query databases, call APIs. But today:

• ❌ Agents have unlimited access to all tools by default
• ❌ No one knows what the agent actually did
• ❌ No difference between read and write operations
• ❌ Multiple agents share the same permission space
• ❌ Enterprise compliance teams can't audit agent behavior

The Solution

from vaultagent import VaultAgent

vault = VaultAgent(api_key="va_sk_xxx")  # or use local policy file

@vault.protect()
def send_email(to: str, subject: str, body: str):
    """This tool requires human approval before execution."""
    return email_client.send(to, subject, body)

@vault.protect()
def query_database(sql: str):
    """This tool is allowed but constrained to SELECT only."""
    return db.execute(sql)

That's it. 3 lines to add permission control. Every tool call is now:

• ✅ Checked against your policy (allow / deny / require approval)
• ✅ Logged with full audit trail
• ✅ Visible in your real-time dashboard
• ✅ Controllable by your team, not just your developers

How It Works

Your AI Agent
    │
    │ tool call: send_email(to="john@...", subject="Refund")
    ▼
┌─────────────────────────────┐
│      VaultAgent SDK         │
│                             │
│  1. Check policy            │  → Is this agent allowed to send email?
│  2. Validate constraints    │  → Is the recipient in the whitelist?
│  3. Apply rate limits       │  → Has the agent exceeded its quota?
│  4. Decision:               │
│     ✅ allow → execute      │
│     ❌ deny → block         │
│     ⏳ require_approval →   │  → Send to Dashboard for human review
│  5. Log everything          │  → Audit trail to Dashboard
└─────────────────────────────┘
    │
    ▼
Actual tool execution (or rejection)

Quick Start

Installation

pip install vaultagent

Option 1: Local Policy (no cloud needed)

from vaultagent import VaultAgent

vault = VaultAgent(policy="vaultagent.policy.yaml")

# vaultagent.policy.yaml
version: "1.0"
agents:
  my-agent:
    tools:
      - tool: "database.query"
        action: allow
        constraints:
          operations: ["SELECT"]
      - tool: "email.send"
        action: require_approval
      - tool: "*"
        action: deny
defaults:
  action: deny

Option 2: Cloud Dashboard

from vaultagent import VaultAgent

# Policy managed in Dashboard, audit logs sent to cloud
vault = VaultAgent(api_key="va_sk_xxx", agent_id="my-agent")

Sign up at vaultagent.dev to get your API key.

Framework Integrations

OpenAI

from vaultagent.middleware import wrap_openai
from openai import OpenAI

client = wrap_openai(OpenAI(), vault)
# All tool calls are now protected by VaultAgent

Anthropic Claude

from vaultagent.middleware import wrap_anthropic
from anthropic import Anthropic

client = wrap_anthropic(Anthropic(), vault)

LangChain

from vaultagent.middleware import wrap_langchain_tools

tools = wrap_langchain_tools([search_tool, email_tool], vault)
agent = create_agent(llm, tools)

MCP (Model Context Protocol)

VaultAgent can run as an MCP proxy server, adding permission control to any MCP server:

vaultagent mcp-proxy \
  --upstream "npx -y @modelcontextprotocol/server-github" \
  --policy vaultagent.policy.yaml \
  --api-key va_sk_xxx

Works with OpenClaw, Claude Desktop, Cursor, and any MCP-compatible client.

OpenClaw

VaultAgent integrates with OpenClaw to add runtime permission enforcement, audit logging, and human approval workflows. No changes to OpenClaw code required — the VaultAgent MCP proxy wraps any upstream MCP server that OpenClaw connects to.

Quick start: Add the VaultAgent MCP proxy to your openclaw.json:

{
  "mcpServers": {
    "filesystem-protected": {
      "command": "node",
      "args": [
        "/path/to/VaultAgent/mcp-server/dist/index.js",
        "--policy", "/path/to/vaultagent.policy.yaml",
        "--agent-id", "openclaw",
        "--", "npx", "-y", "@modelcontextprotocol/server-filesystem", "/workspace"
      ]
    }
  }
}

For deeper integration beyond MCP, install the OpenClaw plugin:

cd openclaw-plugin && npm install && npm run build

See the full OpenClaw Integration Guide for policies, Docker deployment, and troubleshooting.

Dashboard

The VaultAgent Dashboard gives your team real-time visibility into agent behavior:

• 📊 Live Monitor — Real-time feed of every tool call
• 🔒 Policy Manager — Visual editor for permission rules
• ⏳ Approval Queue — Review and approve sensitive operations
• 📋 Compliance Reports — One-click SOC2/HIPAA/GDPR exports
• 🔔 Smart Alerts — Slack/email notifications for anomalies
• 👥 Team Management — SSO + role-based access control

Policy Reference

Actions

Constraints

constraints:
  tables: ["users", "orders"]        # Database table whitelist
  operations: ["SELECT"]             # SQL operation whitelist
  paths: ["/tmp/reports/*"]          # File path whitelist
  domains: ["*.company.com"]         # Email/URL domain whitelist
  max_rows: 10000                    # Result size limit

Rate Limits

rate_limits:
  max_calls_per_minute: 30
  max_calls_per_hour: 500
  max_calls_per_day: 5000

Architecture

┌─────────────────────────────────────────────────┐
│            VaultAgent Cloud Dashboard            │
│  ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐  │
│  │Monitor│ │Policy│ │Approve│ │Report│ │Alert │  │
│  └──────┘ └──────┘ └──────┘ └──────┘ └──────┘  │
└────────────────────┬────────────────────────────┘
                     │ API
    ┌────────────────┼────────────────┐
    ▼                ▼                ▼
┌────────┐    ┌────────┐    ┌────────┐
│ App A  │    │ App B  │    │ App C  │
│  SDK   │    │  SDK   │    │MCP Proxy│
└────────┘    └────────┘    └────────┘

Pricing

Self-hosted SDK and Dashboard are always free and open-source.

Self-Hosting

You can run the entire VaultAgent stack on your own infrastructure — no cloud account needed.

Option 1: Docker (Recommended)

git clone https://github.com/aplomb2/VaultAgent.git
cd VaultAgent

# Start the dashboard
docker compose up -d

# Dashboard is now at http://localhost:3000

Option 2: Manual

# 1. Set up Supabase database
#    Create a project at https://supabase.com, then run:
#    supabase/schema.sql in the SQL Editor

# 2. Configure environment
cd dashboard
cp .env.example .env
# Fill in your Supabase and OAuth credentials in .env

# 3. Start the dashboard
npm install
npm run build
npm start
# → http://localhost:3000

# 4. SDK (Python)
pip install vaultagent

# Point SDK to your self-hosted dashboard
vault = VaultAgent(
    policy="policy.yaml",
    api_key="your-local-key",          # Configure in dashboard
    cloud_endpoint="http://localhost:3000/api/v1"
)

Option 3: SDK Only (No Dashboard)

pip install vaultagent

# Pure local mode — logs to file, no cloud dependency
vault = VaultAgent(policy="policy.yaml", audit_file="audit.jsonl")

All audit logs stay on your machine. Zero data leaves your network.

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

License

Apache 2.0 — see LICENSE for details.

VaultAgent — Because AI agents shouldn't have root access.

Website · Docs · Dashboard · Discord