The era of casually pasting API keys into chat windows is ending. A new tool called TenVault AgentVault, built in Rust and Tauri, introduces a rigorous permission layer between AI agents and sensitive data. Developed by cjdarklord, the solution addresses a critical vulnerability: standard Model Context Protocol (MCP) servers often hold master tokens that grant agents unrestricted access to everything in a vault. This new approach ensures that secrets like Stripe keys or database passwords remain local and are released only upon explicit user approval.

The Problem With Standard MCP Integrations

Most current implementations of MCP for secrets management follow a dangerous pattern: Agent -> MCP Server -> Vault API. In this flow, the MCP server typically possesses full access rights. If the server is compromised or subjected to prompt injection, the agent can exfiltrate every secret the token allows without the user noticing. TenVault AgentVault rejects this 'trust but verify' model in favor of 'deny by default.' The agent does not see the values; it only sees the names of records explicitly shared. The actual secret value is held back until a human clicks 'Allow' in the desktop UI.

Leases, TTLs, and Audit Trails

The core mechanic is the 'lease.' When an agent requests a secret, it doesn't get permanent access. Instead, it receives a temporary grant with a Time-To-Live (TTL), defaulting to one hour. The system limits the number of concurrent leases per record to three, and includes a 'revoke all' button for instant lockdown. Every transaction is logged in an encrypted audit file, derived from the user, machine, and vault path. This creates a forensic trail that shows exactly who asked for what, when, and how the request ended. The master password never leaves the local environment, and the MCP server communicates via a local bridge token, not the master credential.

Technical Implementation and Limitations

Built on a Rust core with XChaCha20-Poly1305 encryption and SQLCipher for the database, the tool runs as a local HTTP server with a Tauri GUI. It includes strict process checks, allowing administrators to whitelist specific parent processes like cursor.exe. However, the author is honest about current limitations: the Windows binary is currently unsigned, triggering SmartScreen warnings, and mobile approvals are still in progress. While Telegram and VK notifications with Allow/Deny buttons are available, the native mobile app is not yet fully integrated. This is a surface reduction strategy, not a silver bullet; if an agent is compromised after receiving a value, it can still use that secret until the lease expires.

Key Takeaways

  • AgentVault uses time-limited leases (TTL) to prevent permanent secret exposure to AI agents.
  • The master password is never passed to the MCP server; access is granted via a local bridge token.
  • An encrypted audit log records every secret request, approval, and denial for forensic review.
  • Current limitations include an unsigned Windows binary and lack of native mobile approval support.

The Bottom Line

This is the security architecture AI agents actually need. We spent the last year teaching bots to code; now we need to teach them to beg for their keys. AgentVault proves that local-first, permission-based access is the only sane way to integrate LLMs with real-world credentials.