ardurai/mara

Mara is a log collection and forwarding agent specifically designed for AI workloads, things like AI coding assistants, local language model runtimes, and automated AI agents. It fills a gap that general-purpose logging tools do not cover well: these AI tools produce telemetry, meaning usage data, token counts, prompt activity, and model calls, in formats that conventional observability systems do not natively understand. The project describes itself as similar in concept to Fluent Bit, a well-known lightweight log shipper, but purpose-built for AI runtimes. It collects telemetry from tools like Claude Code, Codex, Cursor agents, and local language model servers, normalizes the data according to the OpenTelemetry gen_ai semantic conventions, a standard schema for describing AI operations, applies a policy stage that can redact sensitive information before data leaves your machine, and then forwards everything to whatever monitoring backend you already use. Options include Grafana Loki, Splunk, Elasticsearch, Kafka, S3-compatible storage, Prometheus, local files, or a generic webhook. The policy stage runs in a WebAssembly sandbox, meaning plugins that process your data cannot access the rest of your system. Prompt and raw message body capture is opt-in and disabled by default. Mara is written in Rust as a single binary designed to run at the edge, directly on the machine where AI tools are running. It is pre-1.0 and Apache 2.0 licensed. You would use it if you need auditable, structured logs from AI agent activity without sending raw data to a third-party service.