Traccia

license

Apache-2.0(OSI, single root LICENSE)

source

public(the whole tracing product - init/observe, automatic instrumentation, all listed exporters (Grafana Tempo, Jaeger, Zipkin, SigNoz, console, file) - ships and runs with no account)

core-gated

ungated(prompt-registry fetch (load_prompt) and persisted evaluation experiments call the hosted app.traccia.ai platform, but that platform is a separate product sold beside the SDK, not a withheld part of the tracing SDK itself

The tracing, instrumentation and offline-evaluation core is a complete, self-contained Apache-2.0 SDK that exports over OTLP to any collector with no account required. Two named functions - `load_prompt` (a prompt registry) and persisted `evaluate()` experiments - require a workspace API key against the hosted Traccia platform, which has no self-hostable equivalent in this repository. That is the same shape the ladder already resolves as ungated for `pydantic-ai`+Logfire and `langchain`+LangSmith - a complete open core with an optional hosted platform sold beside it, not a paid tier carved out of the core's own functionality. Recorded at medium confidence because, unlike those precedents, the optional platform is Traccia's own commercial product rather than an unrelated one.

• https://raw.githubusercontent.com/traccia-ai/traccia-py/main/LICENSE recorded 2026-09-02

  Apache License, Version 2.0 full text, the sole license file in the repository.

• https://api.github.com/repos/traccia-ai/traccia-py/git/trees/main?recursive=1 recorded 2026-09-02

  The whole repository tree, 140 paths. No ee/, enterprise or premium path exists; the only LICENSE is the root Apache-2.0 file.

• https://raw.githubusercontent.com/traccia-ai/traccia-py/main/README.md recorded 2026-09-02

  The Quick Start and "With LLM Calls" examples run `init()`/`@observe` with no API key and export via OTLP. By contrast, `load_prompt` states "Requires the Traccia platform: create a workspace API key" and `evaluate()`'s persisted-experiment path "Use the same workspace API key as tracing"; `persist=False` is offered for "a local-only loop" of evaluate() without the platform.

The traccia package draws 764 PyPI downloads a month, well inside the <10K band. The package is the scored product - the SDK itself, not a client for the hosted Traccia platform - so its install count is the product's own usage. 108 GitHub stars on a repository created in January 2026 are consistent with a very young product still building adoption.

• https://pypistats.org/api/packages/traccia/recent recorded 2026-09-02

  last_month = 764 downloads for traccia (last_week 128, last_day 5)

• https://pypi.org/pypi/traccia/json recorded 2026-09-02

  info.name traccia, version 0.1.28, project_urls Homepage and Repository https://github.com/traccia-ai/traccia-py, license Apache-2.0, summary "Production-ready distributed tracing SDK for AI agents and LLM applications" - the package points back at the declared repository.

• https://api.github.com/repos/traccia-ai/traccia-py recorded 2026-09-02

  stargazers_count = 108, forks_count = 20, created_at 2026-01-21, pushed_at 2026-08-24

Tracing, prompt management and evaluation-with-experiments together cover every dimension this category grades, and the governance/compliance evidence layer is a genuine addition over the anchor. That puts it level with Langfuse at 4, with the caveat that the prompt and persisted-evaluation halves depend on the hosted platform rather than shipping fully self-contained.

• https://raw.githubusercontent.com/traccia-ai/traccia-py/main/README.md recorded 2026-09-02

  Feature list: automatic instrumentation, framework integrations, LLM-aware tracing (tokens, costs, prompts, completions, latency), guardrail detection, "AI Governance and Compliance: EU AI Act and HIPAA-oriented evidence", prompt management via `load_prompt`/`prefetch_prompts`, and offline evaluation via `evaluate()` running a task and scorers over a dataset.