What is claude-code-proxy?

claude-code-local-observability is a local proxy tool that sits between Claude Code and Anthropic's API, letting a developer see and record every request and response Claude Code makes. It runs on your own machine, intercepting API calls, saving prompts and responses into a local SQLite database, and sending metrics into a local Grafana dashboard stack, without sending that data to any external observability service. To use it, a developer copies an example environment file and edits a routing configuration file called upstreams.yaml, which decides where each request should be sent based on the model name being called. Requests can be routed either straight to Anthropic's API or to a Databricks AI Gateway, useful for teams that have their own internal routing setup. The observability stack, which includes Grafana, Prometheus, and an OpenTelemetry collector, is started with a single make command, and the proxy itself is started separately with a Python command using the uv tool, or through a convenience script that starts both at once. Once running, Claude Code is pointed at the local proxy by setting a base URL in its settings file, and the proxy passes requests through with no added delay by writing the streamed response back to Claude Code while also logging it. Every request is recorded with details like the model used, token counts, and cost, which can be queried directly from the SQLite database with a simple query. Removing the base URL setting reverts Claude Code to talking to Anthropic directly. This tool is aimed at developers or teams who want visibility into their own Claude Code usage, such as tracking costs or reviewing past prompts and responses, without routing that data through a company's existing monitoring pipeline. It requires some comfort with editing YAML configuration and running local services.