papercomputeco/tapes

Transparent telemetry collector for hi-fidelity agent traces

What it solves

Tapes provides a way to capture, store, and analyze the interactions between AI agents and LLMs. It solves the problem of "black box" agentic workflows by providing transparent telemetry, allowing developers to see exactly how an agent reached a conclusion through a structured history of sessions, traces, and spans.

How it works

The system uses an append-only log of raw LLM interactions. A "deriver" process then projects these raw logs into a structured read model consisting of sessions, traces, and spans. This process is idempotent and deterministic, meaning the same input always produces the same structured output.

Internally, it uses content addressing (merkle nodes) for provenance and deduplication. Users can capture data via a proxy and query the results through a dedicated API, with support for semantic search over spans and export capabilities via JSONL.

Who it’s for

Developers building AI agents who need durable storage of agent sessions and a way to instrument and debug their agentic workflows using OpenTelemetry-compatible models.

Highlights

  • Immutable Capture: Every interaction is persisted to an append-only raw log.
  • Deterministic Projection: Raw logs are converted into a browsable session/trace/span model that can be reproduced exactly.
  • OpenTelemetry Instrumentation: Plug-and-play instrumentation for agentic telemetry.
  • Semantic Search: Ability to perform semantic search at the span grain.
  • Local-First Setup: Easy deployment via Docker with bundled PostgreSQL and Ollama support.

Related

  • Project
    monocle2ai/monocle

    An open-source observability and testing framework for GenAI apps and agents that uses OpenTelemetry to provide structured tracing and trace-based assertions.

  • Project
    superloglabs/superlog

    An open-source agentic telemetry system that ingests OpenTelemetry data and groups noisy signals into incidents for easier production debugging.

  • Project
    truera/trulens

    An OpenTelemetry-native evaluation framework for AI agents and RAG apps that uses LLM judges to trace and score application performance, cost, and quality.

  • Project
    lmnr-ai/lmnr

    An open-source observability platform for AI agents that provides tracing, monitoring, and evaluation tools to debug and optimize agentic workflows.

  • Project
    canwhite/AgentEval

    A transparent HTTP proxy that captures Agent-LLM traffic to provide automated grading, behavioral diagnosis, and source-code probing for AI agent evaluation.