AI agents get dummy API keys; the proxy swaps in the real ones

What it does

Agent Vault is a Go-based credential broker that sits between AI agents and the APIs they call. You store real secrets in the vault, configure agents with dummy placeholders like __anthropic_api_key__, and route all outbound traffic through the proxy. It intercepts every request, substitutes the dummy values for live credentials, and forwards the traffic—so the agent itself never possesses anything worth stealing.

The interesting bit

The project treats prompt injection as an inevitability rather than a solvable bug. Rather than hardening the agent, it removes the target entirely by keeping secrets on a separate host and using a MITM proxy architecture to inject auth headers mid-flight. It is essentially a purpose-built, agent-aware forward proxy with a secrets manager attached.

Key highlights

• Replaces dummy header values with real credentials on outbound requests, or rewrites auth headers entirely before they reach the target API.
• Ships as a single binary that runs both the management API/UI and the transparent HTTPS proxy; designed to live on a separate machine from the agent.
• Supports pluggable backends, including Infisical itself or a local encrypted store.
• Provides egress filtering and request logging so you can restrict which endpoints an agent reaches and audit its traffic.
• Works with existing tools transparently by bootstrapping the agent environment via HTTPS_PROXY, requiring no changes to MCP servers, CLIs, or SDKs.

Verdict

Worth evaluating if you run autonomous coding agents or sandboxed LLM tools and want to shrink the blast radius of prompt injection. Less useful if your agents already operate in a fully trusted, non-networked environment where credential exfiltration is not a concern.