Built on AVP Specification · Built on Open Protocol

Your Secrets.
Your Rules.

AI agents have full access to your credentials and no persistent memory. AgentVault encrypts secrets, stores agent memories, and serves both via MCP. One vault. Full control.

Get Started Documentation

~/project

Quickstart

Start in 60 Seconds

$ npm install -g @inflectiv-ai/agentvault

$ agentvault init

$ agentvault wrap -p moderate "claude"

Creates a local encrypted vault, scopes credential access, and runs your agent with only what it needs.

The Problem

Your AI Agent Can See Everything

AI coding agents run as processes on your machine with your full permissions. Every env var, every .env file, every AWS key, visible and accessible.

Full Access by Default

Claude Code, Cursor, Codex. They inherit your entire shell environment. AWS keys, database URLs, API tokens. All of them.

Zero Visibility

No audit trail. No logs. You have no idea which credentials your agent accessed, what it did with them, or when.

Prompt Injection Risk

Dec 2025: 30+ vulnerabilities found across AI coding tools. Agents can be hijacked to exfiltrate your secrets silently.

Features

Built for Security-Conscious Developers

Everything you need to control AI agent credential access.

Encrypted Vault

AES-256-GCM encryption for all stored secrets. Master key derived from your passphrase with scrypt.

Permission Profiles

Restrictive, moderate, permissive, or create custom profiles with granular per-credential rules.

Full Audit Trail

Every credential access logged with agent ID, timestamp, credential name, and allow/deny/redact result.

Kill Switch

One command revokes all active agent sessions. Instantly cuts credential access across all running agents.

Trust Levels

Score agents 1-100. Set minimum trust thresholds per credential. New agents start restricted.

TTL & Expiry

Time-limited sessions. Credentials auto-expire after your configured TTL. No stale access.

Agent Memory

Store knowledge, query caches, and operational data. Keyword search with confidence scoring, freshness decay, and TTL expiry.

MCP Server

Built-in Model Context Protocol server. 12 tools for secrets, memory, audit, and status. Rate-limited and budget-tracked.

Skill Sources

Acquire skills from ClawHub, GitHub repos, or custom registries. Auto-save to encrypted vault memory with content chunking and license enforcement.

How It Works

Four Steps. Full Control.

agentvault init

Initialize

One command creates your vault, encryption keys, and default permission profiles. Takes 2 seconds.

agentvault wrap -p moderate "agent"

Wrap & Sandbox

Run any process, including AI agents, dev servers, and builds, inside a sandboxed environment. Credentials are filtered, redacted, or denied based on your profile.

agentvault memory query "auth flow"

Remember & Learn

Agents store and query encrypted memories. Knowledge persists across sessions with keyword search, confidence scoring, and TTL.

agentvault audit show

Audit & Revoke

See every credential access attempt with timestamps, agent IDs, and results. Kill all sessions instantly.

Comparison

Without AgentVault vs With AgentVault

Without AgentVault

Full .env exposure
No per-process control
No audit logs
No revocation
No secure memory

With AgentVault

Scoped credential access
Encrypted vault
Per-process controls
Full audit logs
Instant revocation
Encrypted memory

Use Cases

Sandbox Anything.

AgentVault isn't just for AI agents. Wrap any process to control exactly which environment variables it can see.

AI Agent Sandboxing

Run Claude, Cursor, or any AI coding agent with scoped credentials. Agents only see what you allow.

$ agentvault wrap -p moderate "claude"

Dev Servers

Start local dev servers with a controlled environment. No accidental credential leaks to hot-reloading processes.

$ agentvault wrap -p moderate "pnpm dev"

CI/CD Pipelines

Build and deploy with only the variables the step needs. Each pipeline stage gets its own scoped profile.

$ agentvault wrap -p ci-deploy "npm run build"

Third-Party Scripts

Run untrusted or third-party scripts without exposing your full environment. Audit everything they touch.

$ agentvault wrap -p restrictive "npx some-tool"

AI Memory Persistence

Store learned patterns, API behaviors, and project context across sessions. Agents query memories before re-discovering known solutions.

$ agentvault memory query "error handling patterns"

MCP Integration

Connect any MCP-compatible agent. Secrets and memories served as tool calls with built-in rate limiting and budget tracking.

$ agentvault mcp start