Large language model (LLM) agents increasingly interact with external tools, shared memory, and orchestration layers, shifting system risk from isolated model outputs to runtime action selection and execution. In such settings, static guardrails and post-hoc evaluation provide limited protection against indirect prompt injection, unsafe tool invocation, and behavioral drift across long-horizon workflows. This paper presents a Governance Proxy, a mediation layer for agentic AI orchestration positioned between agents and Model Context Protocol (MCP) servers. The proxy combines three control mechanisms: (a) capability boundaries that enforce least-privilege access to tools and parameter ranges, (b) a context-aware policy shield that can allow, block, rewrite, or escalate proposed actions, and (c) trust-aware routing that reduces autonomy when repeated violations or unstable behavior are detected. The contribution lies in integrating these mechanisms into a single middleware layer at the language-to-action boundary, rather than relying on model-internal safeguards alone. The paper formalizes the proxy design, describes its request-handling logic, and evaluates it in an MCP-mediated database management setting. The results indicate improved containment of unsafe tool-use behavior with moderate runtime overhead, highlighting the trade-off between agent autonomy and governed execution.