Tapes is a transparent telemetry collector for capturing and querying agent traces from LLM interactions.
The tool solves the problem of durably recording and analyzing agent behavior by implementing an append-only capture system paired with a deterministic projection model. Every LLM interaction is persisted to an immutable raw log, and a pure deriver projects this log into a queryable hierarchy of sessions, traces, and spans with deterministic IDs. This means the same raw input always produces identical derived structures, and re-running the deriver is idempotent. The system stores data in PostgreSQL with pgvector support and exposes a REST API for listing sessions with aggregated metrics, browsing traces and spans, running semantic search over spans, and accessing the original captured interactions verbatim.
Tapes suits teams building AI agents who need to inspect, debug, and analyze agent behavior at scale. It works as a drop-in OpenTelemetry instrumentation layer, capturing interactions transparently without requiring changes to agent code. The tool supports both local development with bundled Docker containers running Postgres and Ollama, and production deployments with external embedding providers like OpenAI. Choose this if you need content-addressed, immutable audit trails of agent sessions with the ability to query and search across past interactions.
The maintainers respond to issues and pull requests within hours. Work in the issue tracker centers on enhancements, bug fixes, and documentation.
