Tapes is a transparent telemetry collector for AI agents, written in Go and designed to capture and organize large language model interactions in a structured, queryable format. The system provides durable storage of agent sessions through an append-only capture mechanism that persists every intercepted LLM interaction to an immutable raw_turns log. This foundational design ensures that all agent work is permanently recorded and can be audited or replayed at any time.
The core architecture separates capture from querying through a pure, idempotent deriver that projects the raw interaction log into a hierarchical read model organized as sessions, traces, and spans with span links. This separation of concerns means that re-running the deriver simply reproduces the same projection, and the derived IDs are deterministic so identical raw inputs always yield identical sessions, traces, and spans. The system uses content addressing internally for provenance and deduplication, though this merkle node layer remains hidden from user-facing interfaces.
Tapes provides multiple query surfaces for accessing captured data. Users can list and inspect sessions through paginated endpoints with aggregations for model, token, cost, and turn counts. The system supports browsing traces and spans, running aggregations at the span grain, and performing semantic search across captured spans. The original raw capture is always available verbatim through dedicated endpoints, ensuring complete transparency and auditability of all agent interactions.
The system integrates with PostgreSQL for persistent storage and requires the pgvector extension to support semantic search capabilities. For embeddings, Tapes supports multiple providers including local embeddings through Ollama and cloud-based embeddings through OpenAI. The quickstart documentation indicates that users can bootstrap a local development environment using Docker, which bundles both PostgreSQL and Ollama for immediate experimentation.
According to GitGenius activity tracking, the repository shows strong engagement with a median issue and pull request response latency of 0.1 hours and a mean latency of 9.8 hours across 32 tracked items. The most active contributor, bdougie, has generated 63 events, followed by jpmcb with 23 events. Enhancement requests represent the most common issue type with 7 tracked items, followed by bug reports with 4 items and documentation issues with 2 items. The repository maintains connections with related projects including modelcontextprotocol/inspector, aaif-goose/goose, and block/goose through overlapping contributors.
Tapes is positioned within the broader ecosystem of agentic AI tooling, with topics including agentops, AI, OpenTelemetry, and telemetry. The system supports exporting captured conversations in multiple formats including Markdown transcripts and JSONL, and provides a terminal user interface deck for browsing sessions and drilling into individual session details. The project maintains documentation at tapes.dev and actively solicits community contributions through its contributing guidelines.
