keyblind

Keyblind is an encryption key vault designed specifically for AI agents, with the core idea of "hide your keys from AI." It integrates with AI tools through MCP (Model Context Protocol), ensuring keys are resolved at runtime while plaintext values never appear in LLM conversation records.

Purpose

When using AI programming tools, developers often inadvertently leak API keys, passwords, and tokens to AI. According to available information, in 2025 more than 100,000 LLM conversations containing exposed keys were indexed by search engines. AI agents read .env files, copy and paste keys into conversations, or accidentally commit keys. Keyblind prevents these leaks through encryption at rest and runtime resolution.

Main Capabilities

• MCP-first: Works with all AI tools that support the Model Context Protocol, including Claude Code, Cursor, Copilot, Windsurf, Cline, and Zed.
• Runtime resolution: AI agents can request the use of a key (such as "use my OPENAI_API_KEY"), but the key value does not appear in the conversation record.
• Automatic configuration: Through keyblind setup-mcp command automatically configures Claude Code and can be run in any directory.
• Biometric access control: Optional Touch ID verification requires biometric authentication before resolving keys, and sessions expire after 15 minutes (requires a Pro or Team license).
• Browser management interface: Manage keys through app.keyblind.dev, supporting viewing, adding, copying, and deleting keys, as well as viewing audit logs and license management.
• Chrome extension: Detects and prevents keys from being pasted into AI chat interfaces (Claude.ai, ChatGPT, Copilot), can recognize more than 12 API key formats (OpenAI, GitHub, Stripe, AWS, etc.), and displays a warning banner when a key is detected.
• Multiple key backends: Supports multiple key storage backends.

Usage

1. Install and run keyblind setup-mcp to automatically configure Claude Code, then restart Claude Code.
2. For other editors, refer to the editor-specific configuration, or add .mcp.jsonin the project root directory, or use claude mcp add.
3. After configuration, you can directly tell the AI "list my keyblind secrets" or "use my OPENAI_API_KEY", and the AI agent will resolve the key at runtime without seeing the plaintext.
4. If biometric access control is needed, Touch ID must be configured and a Pro or Team license is required.
5. Browser management: start the HTTP server, then log in at app.keyblind.dev/login using the license key.
6. Chrome extension: located in the browser-extension/ directory, load it from chrome://extensions as an unpacked extension.

Security Features

• AES-256-GCM encryption, PBKDF2 key derivation (600,000 iterations).
• Machine identity-bound key: the encryption key is XOR-wrapped with the machine fingerprint.
• Zero network, zero telemetry: no cloud, no account, no analytics.
• The vault is stored in ~/.keyblind/, with permissions set to 0700.
• Uses HMAC-SHA256 to generate deterministic sandbox fake values for each project and key name.

Use Cases

Suitable for developers using AI programming tools (such as Claude Code, Cursor, etc.) who want to prevent API keys, passwords, and tokens from leaking in AI conversations. Especially suitable for teams and individuals who need to securely manage multiple keys, need audit logs, or need additional biometric protection.

Limitations

• The biometric access control feature requires a Pro or Team license; specific pricing and how to obtain it are not yet provided.
• The browser management interface requires starting the HTTP server and logging in with a license key.
• The Chrome extension needs to be manually loaded as an unpacked extension.
• The specific configuration methods for other editors need to refer to the editor-specific configuration documentation.
• Details such as the list of supported specific key backends, system requirements, and installation prerequisites are not fully provided in the materials.