Agent Vault

Agent Vault is an open-source credential broker from Infisical that sits between AI agents and the APIs they call, so agents never hold real credentials. It targets credential exfiltration: an agent that holds a key can be tricked through prompt injection into leaking it. Instead, you store keys such as an Anthropic API key or a GitHub token in Agent Vault, route the agent's HTTP traffic through it, and Agent Vault attaches the credentials to each outbound request before forwarding it.

It ships as a single binary that acts as both server and CLI client, combining a vault, a management UI and API, and a man-in-the-middle (MITM) HTTPS proxy. Agents keep dummy placeholder values in their own environment, and service rules tell the proxy to swap those for the real secrets or replace auth headers entirely. Because the agent's environment is bootstrapped to use the proxy, existing MCP servers, CLIs, SDKs, and API calls work unchanged. Agent Vault is meant to run on a separate host from the agents, on the same private network.

Each long-running agent gets its own token. For ephemeral sandboxes, an orchestrator mints a short-lived, vault-scoped token through the TypeScript SDK and passes the proxy settings and CA certificate into the sandbox, whether that is Docker, Daytona, E2B, or Firecracker. Use cases include remote Claude Code sessions that open pull requests, all-purpose agents such as OpenClaw and Hermes, and custom harnesses.

State lives in SQLite by default, with PostgreSQL for multi-instance deployments. A vault can also be backed by an external store such as Infisical, which adds dynamic secrets. A separate commercial Infisical Agent Vault, built into the Infisical platform, adds access bundles, time-bound sessions, and Infisical access controls. The project is in preview.