AI agents need access to credentials — but they should never see them.Every time you paste a password, API key, or token into @NousResearch Hermes, @openclaw Openclaw, @claudeai Claude Code, @OpenAI Codex, @xai Grok Build, @GeminiApp Gemini, or any other AI tool, you're sending plaintext through the context window. That data gets cached, logged, and potentially used for training.
This has become one of the biggest security concerns for engineers building and using AI agents in daily workflows.
That’s exactly why I created psamvault-cli and its companion psamvault-mcp.
How it works:
- When you save a credential with the psamvault CLI, it’s encrypted immediately on your device — before it ever leaves your machine.
- The encrypted blob is stored on the server, but we never have the decryption key.
- Your decryption key stays securely on your device, protected by your OS keychain (iCloud Keychain, Windows Credential Manager, or Linux Secret Service).
When an AI agent needs to use a stored credential, psamvault-mcp handles it seamlessly: the agent can authenticate and perform tasks without ever seeing the actual password or key.
Demo: Here’s me using the @jack Goose agent
to make an authenticated API request with a key stored in the vault (no credentials exposed to the agent).
You can install and try it out via
CLI: psam-717/psamvault-cli
MCP: psam-717/psamvault-mcp
@Teknium@steipete I think I have a solution one major problem using AI agents