impact-boundary-labs/mcpboundary

MCP Boundary is a local proxy that sits between an AI agent and the MCP (Model Context Protocol) servers it uses for tools like email, file access, or databases. When an AI agent calls a tool, the request passes through MCP Boundary first, which checks it against a policy you define before deciding whether to allow, block, or hold the call. If it blocks the call, it sends back a structured reason the agent can read and act on, such as a directive to narrow the scope or re-check current state before retrying. MCP is a protocol used by AI coding tools like Claude and Cursor to connect agents to external services. MCP Boundary wraps any local stdio-based MCP server, meaning the agent's view of available tools stays the same while every call is intercepted for policy checking. You can choose which tools are visible to the agent, which are hidden, and which only appear in the local dashboard for inspection. Beyond simple allow or block rules, MCP Boundary can restrict specific arguments rather than entire tools. For example, you can allow an email tool but only permit it to send to certain recipients. You can also set limits on response sizes, call timeouts, and the frequency of repeated calls to prevent runaway loops. A state-binding feature checks whether conditions have changed since the agent last read data before allowing a write operation. A local dashboard shows every call, the policy decision, the reason, and the outcome. The tool is designed for local command-based MCP servers and is explicitly not an enterprise security product, data-loss prevention system, or prompt-injection detector. Pre-built binaries for Windows and Linux are available on the project website. The code is free to download and use, but all rights are reserved and it ships under a custom proprietary license.