Cloak, from Cloakward, is an open-source Model Context Protocol server that provides a local security layer to keep API keys and credentials out of Large Language Models' view. It functions as a local vault that responds to agent tool calls with secure references instead of plaintext secrets, enabling authenticated actions without exposing raw values. The design emphasizes local-first privacy and protocol compatibility, targeting AI developers and security-conscious power users building agent workflows.

It integrates with the emerging MCP client ecosystem

Cloak runs as an MCP-native server compatible with MCP-aware clients such as Claude Desktop and Cursor, using a standardized protocol to avoid vendor lock-in. Distribution is handled as a Rust-based binary that starts a compact local server process. This approach lets agent platforms request and receive credential references over MCP plumbing instead of adding provider-specific extensions to each client.

It enforces a reference-only flow so models do not receive raw secrets

Rather than returning plaintext, Cloak supplies opaque references and performs authenticated calls on behalf of agents, implementing a zero-exposure credential model. The language model does not receive raw key text during a tool invocation, and the sensitive string does not leave the user's machine in plaintext form. Security depends on correct MCP client configuration and deploying the server locally to maintain the reference-only path.

Input types are explicit and the scope is agent-focused

The server accepts arbitrary sensitive strings with documented examples such as OpenAI API keys, gateway credentials, cloud provider tokens, and GitHub personal access tokens. Supported secret types include:

• OpenAI keys
• Gateway credentials
• Cloud tokens
• GitHub access tokens

Developer workflow favors CLI control and local auditability, with integration limits

Management is command-line oriented, with operations like 'cloak add' for storing secrets and 'cloak remove' for cleanup, and the open-source codebase provides a local audit path. The Rust implementation aims for performance and memory safety. The tool is most applicable where agent platforms already support MCP; teams without MCP-compatible clients cannot use the reference-only secret flow.

Practical judgement: best for MCP-based developer workflows

Cloak is a pragmatic option for AI developers and security-conscious power users who must keep agent-held credentials inaccessible to language models while retaining local control. It assumes teams can deploy a local MCP server and integrate agent clients into that pipeline, so it fits projects willing to operate local services and perform code audits. For environments without MCP support, the tool is not applicable.