Agentsecrets

AgentSecrets is open-source, zero-knowledge credential infrastructure designed for the AI era.

Traditional secrets managers (like Doppler, Infisical, or HashiCorp Vault) inject plaintext credentials into environment variables or process memory. When AI coding assistants (such as Cursor or Claude Desktop) read workspace files or autonomous agents are subjected to prompt injection attacks, those raw API keys and database tokens are easily exfiltrated.

AgentSecrets fundamentally eliminates this attack surface by decoupling credential values from agent execution:

1. Zero-Knowledge Transport Proxy: AI agents and workflows reference credentials strictly by name (e.g. GITHUB_TOKEN). Outbound HTTP calls are intercepted by a local loopback proxy that resolves the encrypted key from your secure OS keychain and injects it into the HTTP header at the transport layer. The agent receives only the upstream API response—raw secret values never enter agent memory, context, or disk.
2. Anti-Impersonation Keychain Daemon: OS keychain access is cryptographically bound to the caller's true PID and binary SHA-256 hash, preventing unauthorized background scripts or malware from accessing stored credentials.
3. Client-Side NaCl SealedBox Encryption: Cloud synchronization across teams and environments uses Curve25519 public-key encryption before leaving your machine. The backend sync server stores only ciphertext it structurally cannot decrypt.
4. Native Model Context Protocol (ZK-MCP): Built-in zero-knowledge MCP server integration for Claude Desktop and Cursor, enabling autonomous agent tooling without credential exposure.
5. Deterministic Audit Chains: Every access attempt is recorded in a cryptographically verifiable SHA-256 hash chain with no value field to leak.