登录

logfire

pydanticMIT活跃维护

面向生产环境 LLM 与 Agent 系统的可观测平台，基于 OpenTelemetry 采集日志、指标和调用链，支持评估与调试。

AI observability platform for production LLM and agent systems.

v5.5.0发布于 2026-11-07 00:10 UTC4,544star +7 / 7 天361fork同步于 2026-11-10 11:37 UTC
LLM 可观测Agent 监控OpenTelemetry调用链追踪评估
对比 GitHub 官网agent-observabilityaiai-observabilityai-toolsevals

怎么上手

由 README 提炼，执行前请确认
原文 
pip install logfire
uvx logfire-cli auth
uvx logfire-cli init
硬件要求按 README 说明

README 未说明，纯 Python SDK 无 GPU 要求

下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼
没人工点评？写下你的经验

Pydantic 团队出品的 Python 可观测性 SDK，基于 OpenTelemetry 追踪 LLM 与 Agent 系统

面向生产环境 LLM 与 Agent 系统的可观测平台，基于 OpenTelemetry 采集日志、指标和调用链，支持评估与调试。

核心能力

  • 用 logfire.span 和 logfire.info 手动埋点追踪
  • 一行 instrument_fastapi 自动插桩 FastAPI 等流行包
  • 基于 OpenTelemetry 封装，可导出到任意 OTel 后端
  • 用标准 SQL 查询追踪数据，兼容现有 BI 工具
  • 内置 Pydantic 模型校验数据分析

容易踩的坑

服务端 UI 与后端闭源，自托管需购买企业许可；logfire.configure() 不会引导登录，需先跑 auth 和 init 保存写 token；uvx 命令依赖 uv，未安装 uv 需先装

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

5 个版本v5.5.0
全部 release
  1. v5.5.0最新正式版2026-11-07 · 2026-11-07 00:10 UTC
    • Drop all incoming context when distributed tracing is disabled by @dmontagu in #2679
    • Carry variable overrides and targeting context into thread pools by @dmontagu in #2680
    • Name the setting when LOGFIRE_TRACE_SAMPLE_RATE is not a number by @chrikrah in #2685
    • Record audio responses and legacy function calls from OpenAI chat completions by @pcbeingused333 in #2636
    • Avoid duplicate console output from Loguru by @imp-joshi in #2690
    • Save a read token on Windows, where os.fchmod does not exist by @PRINCE2-AI in #2696
    • Record the explanation of an Anthropic refusal as output text by @pcbeingused333 in #2638
    • Record custom tool calls from the OpenAI Responses API by @michaelkillgta in #2694
    • Scrub AWS security tokens (x-amz-security-token, aws_security_token) by default by @hsusul in #2692
    • Let a span status set by the exception callback stand by @pedro-teixeira in #2707
    • Record legacy function calls in OpenAI chat completion input messages by @chrikrah in #2711
    • Return Logfire instance from logfire-api configure() when SDK is not installed by @PatrickRibbsaeter in #2716
    • fix(datasets): serialize dictionary-valued case fields by @shuaiyafei in #2713
    • Convert Anthropic SDK content blocks sent back as input messages by @jaeminbird in #2718
    • Keep streamed LLM responses iterable when recording them fails by @jaeminbird in #2721
    • Skip empty items in comma-separated Pydantic plugin include/exclude settings by @Fahmid-Arman in #2720
    • Fall back to OpenTelemetry instrumentors in logfire run by @mekanbaymyradov in #2609
    • Breaking Change: Require project credentials before SDK configuration by @strawgate in #2683
    • Connect agent sandboxes over gRPC with pydantic-monty 1.5.0 by @dmontagu in #2737
    • Make SpanLevel compare equal to 'warning' by @Fahmid-Arman in #2732
    • Don't crash the console exporter on attributes that format to an empty string by @Fahmid-Arman in #2731
    • Add configure(fleet=...) to manage the SDK from Logfire Fleet by @strawgate in #2736
    • Add actionable Agent Sandboxes error hints by @dmontagu in #2739
    • Send telemetry with the API key when no write token is configured by @dmontagu in #2743

    New Contributors

    • @pierrebrunelle made their first contribution in #2644
    • @PRINCE2-AI made their first contribution in #2696
    • @michaelkillgta made their first contribution in #2694
    • @hsusul made their first contribution in #2692
    • @pedro-teixeira made their first contribution in #2707
    • @PatrickRibbsaeter made their first contribution in #2716
    • @shuaiyafei mad
  2. v5.4.02026-10-26 · 2026-10-26 15:07 UTC
    查看更新说明
    • logfire.agent_sandboxes() by @samuelcolvin in #2658 and #2672
    • fix(sdk): serialize all pathlib.Path variants as plain strings by @RichardoMrMu in #2669
    • Format an empty set as set() in the console exporter by @chrikrah in #2671
    • Keep binary attributes out of the secret scrubber by @Eapp1e in #2668
    • Keep recording exceptions whose __str__ raises by @Eapp1e in #2667
    • Print generator attributes as their repr in the console exporter by @CyberNativeAI in #2657
    • Record OpenAI cache write tokens on spans by @pcbeingused333 in #2637
    • Keep non-string messages from breaking exporters by @oxura in #2655
    • Count only top-level CTE names in SQL summaries by @gokay-ai in #2652
    • Log dataclass types as their repr, not their default field values by @chrikrah in #2649
    • Record max_completion_tokens as the requested max tokens on OpenAI chat spans by @pcbeingused333 in #2646
    • Record custom tool calls from OpenAI chat completions by @pcbeingused333 in #2647
    • Convert Anthropic document and file-backed image input blocks to semconv parts by @pcbeingused333 in #2645
    • Harden remote strings in log messages by @dmontagu in #2654
    • Remove the duplicated tool call from message content by @RachelWanggg in #2442
    • Record errors raised during LLM stream consumption by @deepanshupal in #2491
    • fix(formatter): treat unbalanced braces as a formatting failure by @Eapp1e in #2675
  3. v5.3.02026-10-23 · 2026-10-23 10:37 UTC
    查看更新说明
    • Automatically instrument FastAPI with logfire run by @alexmojaki in #2606
    • Automatically instrument Starlette apps with logfire run by @alexmojaki in #2596
    • Automatically instrument Flask apps with logfire run by @alexmojaki in #2594
    • Raise when SQLAlchemy instrumentation is skipped by @maaz-bin-hassan in #2602
    • Add pending_spans option by @adriangb in #2580
    • Honor exception callback levels when setting span status by @alexmojaki in #2592
    • Scrub passphrase, signing_key and hmac_key by default by @sridhar-3009 in #2593
    • Scrub OAuth-style token names (access_token, refresh_token, id_token, bearer_token) by default by @sridhar-3009 in #2551
    • Treat logfire.tags passed through structlog as log tags by @Thanatos9404 in #2598
    • Treat logfire.tags passed through logging and loguru as log tags by @Thanatos9404 in #2550
    • Suppress native FastAPI telemetry by @alexmojaki in #2629
    • Require certifi 2026.10.19 to distrust compromised eMudhra chains by @alexmojaki in #2640
    • Format range attributes in the console exporter by @fhgffy in #2534
    • Format bytearray attributes in the console exporter by @sebxt in #2591
    • Fix DataFrame labels below pandas display limits by @fallintoplace in #2509
    • Handle zero-dimensional NumPy arrays in console formatting by @dailyun in #2621
    • Clarify CLI account switching and logout scope by @alexmojaki in #2552
    • Include the dataset in url_from_eval links when the report has one by @jirikuncar in #2566
    • Keep the parent context for spans started while instrumentation is suppressed by @adtyavrdhn in #2570
    • Record response id and finish reasons for streamed OpenAI and Anthropic calls by @flcrom in #2545
    • Record the response model for streamed OpenAI and Anthropic calls by @darshjoshi in #2558
    • Record OpenAI refusals in gen_ai.output.messages by @pcbeingused333 in #2582
    • Record streamed OpenAI Responses that end as incomplete or failed by @SergeevDmitry in #2565
    • Record response id and finish reason on Claude Agent SDK chat spans by @pcbeingused333 in #2620
    • Record cache token usage on Anthropic spans by @pcbeingused333 in #2616
    • Record cached prompt tokens on OpenAI spans by @pcbeingused333 in #2617
    • Record why a Claude Agent SDK run ended in error by @pcbeingused333 in #2622
    • Record finish reasons for OpenAI Responses API calls by @pcbeingused333 in #2619
    • Record response id and finish reasons on OpenAI Agents Responses spans by @pcbeingused333 in #2632
    • Record Responses API refusals as text in gen_ai.output.messages by @pcbeingused333 in [#2618](https://github
  4. v5.2.02026-10-13 · 2026-10-13 17:26 UTC
    查看更新说明
    • Support OpenTelemetry 1.45 by @alexmojaki in #2536
    • Allow console attributes without verbose output by @rudrakshkarpe in #2488
    • Add compression option and zstd experimental support by @Viicos in #2535
    • Update FastAPI instrumentation to prevent duplicate telemetry by @alexmojaki in #2524
    • Update Anthropic instrumentation by @alexmojaki in #2521
    • Keep span message in sync with the name after update_name by @darshjoshi in #2514
    • Expose RFC 9457 problem details on query client errors by @adriangb in #2511
    • Read stable HTTP semconv attributes when building HTTP span messages by @samuelcolvin in #2516
    • Update Google GenAI instrumentation by @alexmojaki in #2523
    • Convert span attribute ints below the int64 minimum to strings by @JoeyTan21 in #2502
    • Respect repr=False when logging attrs fields by @harbinresearcher in #2518
    • Default the CLI base URL to LOGFIRE_BASE_URL by @jgjoe in #2260
    • Create subprocess with no console windows on windows by @marcelotrevisani in #2507
    • Fix when logfire tries to access a file that has no permission by @marcelotrevisani in #2464
    • fix: do not treat non-retryable DiskRetryer responses as success by @alexmojaki in #2529
    • fix explicit context for logfire metrics by @alexmojaki in #2532
    • Fix OTLP byte POST retries through Session.request by @SAY-5 in #2503

    (These are the same changes as in v6.0.0b8, but backported to the v5 line.)

  5. v6.0.0b8pre-release2026-10-10 · 2026-10-10 01:01 UTC
    查看更新说明
    • Support OpenTelemetry 1.45 by @alexmojaki in #2536
    • Allow console attributes without verbose output by @rudrakshkarpe in #2488
    • Add compression option and zstd experimental support by @Viicos in #2535
    • Update FastAPI instrumentation to prevent duplicate telemetry by @alexmojaki in #2524
    • Update Anthropic instrumentation by @alexmojaki in #2521
    • Keep span message in sync with the name after update_name by @darshjoshi in #2514
    • Expose RFC 9457 problem details on query client errors by @adriangb in #2511
    • Read stable HTTP semconv attributes when building HTTP span messages by @samuelcolvin in #2516
    • Update Google GenAI instrumentation by @alexmojaki in #2523
    • Convert span attribute ints below the int64 minimum to strings by @JoeyTan21 in #2502
    • Respect repr=False when logging attrs fields by @harbinresearcher in #2518
    • Default the CLI base URL to LOGFIRE_BASE_URL by @jgjoe in #2260
    • Create subprocess with no console windows on windows by @marcelotrevisani in #2507
    • Fix when logfire tries to access a file that has no permission by @marcelotrevisani in #2464
    • fix: do not treat non-retryable DiskRetryer responses as success by @alexmojaki in #2529
    • fix explicit context for logfire metrics by @alexmojaki in #2532
    • Fix OTLP byte POST retries through Session.request by @SAY-5 in #2503
    • Handle project setup for accounts without organizations by @strawgate in #2469

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    Pydantic Logfire: Know more. Build faster.

    From the team behind Pydantic Validation, Pydantic Logfire is an observability platform built on the same belief as our open-source library: that the most powerful tools can be easy to use.

    What sets Logfire apart:

    • Simple and Powerful: Logfire's dashboard is simple relative to the power it provides, ensuring your entire engineering team will actually use it.
    • Python-centric Insights: From rich display of Python objects, to event-loop telemetry, to profiling Python code and database queries, Logfire gives you unparalleled visibility into your Python application's behavior.
    • SQL: Query your data using standard SQL: all the control and (for many) nothing new to learn. Using SQL also means you can query your data with existing BI tools and database querying libraries.
    • OpenTelemetry: Logfire is an opinionated wrapper around OpenTelemetry, allowing you to leverage existing tooling, infrastructure, and instrumentation for many common Python packages, and enabling support for virtually any language. We offer full support for all OpenTelemetry signals (traces, metrics and logs).
    • Pydantic Integration: Understand the data flowing through your Pydantic Validation models and get built-in analytics on validations.

    See the documentation for more information.

    Evaluating observability tools? See what Logfire does for LLM apps and agents, evals in production, and how it compares to alternatives.

    Feel free to report issues and ask any questions about Logfire in this repository!

    This repo contains the Python SDK for logfire and documentation; the server application for recording and displaying data is closed source.

    Using Logfire

    This is a very brief overview of how to use Logfire, the documentation has much more detail.

    Install

    pip install logfire

    This installs both the Python SDK and the Logfire CLI. To install only the SDK, use pip install logfire-sdk; the import remains import logfire.

    (learn more)

    Connect your project

    uvx logfire-cli auth
uvx logfire-cli init

    Run these commands from your application's directory. uvx is part of uv. Authentication signs you in; init selects or creates a project and saves its write token locally for the SDK. Existing project credentials continue to work. logfire.configure() does not prompt for setup.

    (learn more)

    Manual tracing

    Here's a simple manual tracing (aka logging) example:

    from datetime import date

import logfire

logfire.configure()
logfire.info('Hello, {name}!', name='world')

with logfire.span('Asking the user their {question}', question='age'):
    user_input = input('How old are you [YYYY-mm-dd]? ')
    dob = date.fromisoformat(user_input)
    logfire.debug('{dob=} {age=!r}', dob=dob, age=date.today() - dob)

    (learn more)

    Integration

    Or you can also avoid manual instrumentation and instead integrate with lots of popular packages, here's an example of integrating with FastAPI:

    from fastapi import FastAPI
from pydantic import BaseModel

import logfire

app = FastAPI()

logfire.configure()
logfire.instrument_fastapi(app)
# next, instrument your database connector, http library etc. and add the logging handler


class User(BaseModel):
    name: str
    country_code: str


@app.post('/')
async def add_user(user: User):
    # we would store the user here
    return {'message': f'{user.name} added'}

    (learn more)

    Logfire gives you a view into how your code is running like this:

    Logfire screenshot

    Contributing

    We'd love anyone interested to contribute to the Logfire SDK and documentation, see the contributing guide.

    Reporting a Security Vulnerability

    See our security policy.

    Logfire Open-Source and Closed-Source Boundaries

    The Logfire SDKs (we also have them for TypeScript and Rust) are open source, and you can use them to export data to any OTel-compatible backend.

    The Logfire platform (the UI and backend) is closed source. You can self-host it by purchasing an enterprise license.

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    langfuse
    langfuse·TypeScript

    开源 LLM 应用可观测与评估平台，能追踪调用链路、跑评测实验，帮团队定位问题并持续优化提示词和模型效果。

    LLM 可观测+4LLM 可观测
    37251+1274.2kv4.77.02026-11-10 11:19 UTC
    signoz
    SigNoz·TypeScript

    基于 OpenTelemetry 的开源可观测性平台，把日志、指标和链路追踪集中到一处，提供 APM、分布式追踪和基础设施监控，帮团队快速定位线上问题。

    可观测性+4LLM 可观测
    32568+242.6kv0.149.02026-11-10 10:28 UTC
    skywalking
    apache·Java

    分布式系统的应用性能监控平台，支持链路追踪、指标采集和日志分析，帮你在微服务架构下快速定位性能瓶颈。

    应用性能监控+4LLM 可观测
    24998+46.6kv11.1.02026-11-10 09:54 UTC
    jaeger
    jaegertracing·Go

    CNCF 出品的分布式追踪平台，采集服务调用链路并做可视化分析，帮你在微服务和 LLM 应用里定位延迟与故障根因。

    分布式追踪+4LLM 可观测
    23393+73.2kv2.23.02026-11-10 06:28 UTC