Surfaces read-only telemetry from AgentGuard47, the Python SDK that kills runaway agent runs before they drain your budget. Connect this to Claude Desktop and you can query live traces, check budget health, review loop or retry alerts, and pull cost breakdowns without leaving the conversation. Useful when you're running Python agents that call tools, retry flaky APIs, or review code autonomously and you want visibility into guardrails that actually stopped execution. The MCP server exposes what the in-process guards already wrote to local JSONL traces, so you get incident reports and spend tracking through stdio without adding network calls or external dependencies. Pair it with the core SDK to turn agent forensics into a context you can chat with.
claude mcp add --transport stdio io.github.bmdhodl-agentguard47 uvx agentguard47