MCP+ is an open-source post-processing layer built by Salesforce AI Research that wraps MCP clients to reduce token bloat from tool responses. It intercepts tool outputs and uses a lightweight LLM to extract only the information an agent actually needs, cutting token usage by up to 75% without modifying existing tools or agent code. The system works by injecting an `expected_info` parameter into tool schemas, which agents naturally populate. When a tool response exceeds a configurable token threshold, MCP+ routes it through dual extraction methods — natural language summarization and generated Python code — before returning a concise result. A SafeCodeExecutor sandboxes the generated code to prevent dangerous operations. Setup requires only a CLI command pointing at an existing MCP config file, with no application code changes needed.
