GeminiLens: Observability for Gemini Agents

GeminiLens is a small open-source library plus dashboard that gives Gemini agent builders the four numbers they care about during development: cost per call, p95 latency, output drift over time, and a full audit log of every external host an agent's tools reached. It wraps a Vertex AI Gemini client without forcing you onto a hosted backend. A Streamlit dashboard renders the traces locally, and a one-line exporter pushes them to Dynatrace or TrueFoundry for production monitoring. Why we built it: every Gemini agent project goes through the same loop. You ship the agent, then you spend a week wiring per-call cost, then another week building an "is the output getting worse" check, then a month wishing you had logged the URLs your tools actually hit. GeminiLens collapses those three things into one library plus one dashboard, so you can keep building the agent instead of building observability around it. Built on google-genai for Vertex AI Gemini 2.5 calls, httpx for the egress allowlist, Streamlit and pandas for the dashboard. The cost table is hand-curated from Google's published Gemini pricing and is checked into the repo so reviewers can see what we charge against. Math is pure stdlib, so nothing has to be mocked in tests. The repo ships with an example agent, a recorded 2-minute demo, a one-command Cloud Run deploy, and exporters for Dynatrace and TrueFoundry. Apache-2.0.