MLflow 3.12.0 introduces support for routing Claude Code through the MLflow AI Gateway, giving teams full observability, budget controls, and content guardrails over autonomous coding agent sessions. Setup requires just two environment variables — no code changes needed. Every LLM call is captured as a traceable event in the MLflow UI, showing prompts, responses, token counts, and latency. Budget policies allow setting spending thresholds with alerts and hard limits to prevent runaway costs. Guardrails can screen requests for prompt injection, PII, or restricted topics before they reach Anthropic, and filter responses before the agent sees them. The same approach also works with OpenAI Codex and Gemini CLI.
Table of contents
How It Works Setting Up the Integration Observability: Every Request as a Trace Budget Controls: Keep Spending in Check Guardrails: Enforce Policies Across Every Session Getting Started 268 Impressions