AgentLedger

AgentLedger is an open-source observability and control tool for AI agents. It is built for monitoring what agents actually do, with a focus on actions, costs, and stopping agents when things go wrong. The page frames it around agents that send emails, create tickets, charge credit cards, and call APIs.

It logs every action and tracks every cost. Features listed on the site include a live streaming feed with real-time SSE updates, policy engine templates for rules such as rate limits, allowlists, and cost caps, human approval for high-risk actions, advanced analytics with multi-day trend analysis and cost forecasting, trace replay for step-by-step debugging, and batch logging and data export. The comparison section also mentions statistical anomaly detection, agent evaluations and scoring, budget controls and enforcement, and slack, Discord, and PagerDuty alerts.

AgentLedger provides Python and TypeScript SDKs, a CLI tool, and integrations with LangChain, OpenAI, MCP, CrewAI, AutoGen, LlamaIndex, Vercel AI SDK, and Express. It shows an example of tracking an action with a trace ID and capturing output, and says setup takes three steps and about five minutes with no infrastructure or config files. The site also says it is self-hostable, runs on your own infrastructure, and is fail-open. The pricing section lists a free plan with 5,000 actions per month, 5 agents, and 7-day data retention, plus Pro and Team plans with higher limits and longer retention. The page identifies the codebase as open source and fully auditable on GitHub.