0xkaz/llm-governance-dashboard

This is a self-hostable dashboard for teams that have deployed AI coding tools like Claude Code, Codex, or Cursor and now need visibility into how much money is being spent and whether the tools are actually being used. The system works by placing a proxy gateway in front of every AI provider your team uses. Instead of developers calling OpenAI or Anthropic directly, they point their tools at this local proxy using a per-user API key. Every request passes through LiteLLM, a library that handles routing to the right upstream provider, and each call is recorded as a row in a BigQuery table with details like the user, team, model, tokens used, cost in dollars, and response latency. No personal data is stored, only anonymized identifiers. A FastAPI dashboard turns that log data into two things: cost governance and adoption tracking. The cost side shows spending per user and team, tracks each group against a monthly budget, and sends Slack or email alerts when a team hits 80 or 100 percent of its limit. Virtual API keys can be issued and revoked from the dashboard. The adoption side shows whether people are actually returning to use the tools each week, how many different tools each person uses, and trend charts for daily active usage. It distinguishes between a productivity claim and a simple activation signal. The whole stack runs locally via Docker. You configure a BigQuery project, run a setup command to create the table, then bring up the proxy with Docker Compose. Adding a new provider requires only an entry in a YAML config file and the provider API key. The dashboard is available in English and Japanese. The project is described as a demo and reference implementation, not a production-hardened system.