Skip to content
unnudged
brandRising

Langfuse on X

Langfuse talk is mostly praise: 43 of 76 people in reception recommend it, mainly for tracing, cost and evals in one place, and 6 choose it for being open source and self-hostable. Complaints and alternatives stay small and scattered.

First written 27 Oct · what changed · from 95 posts by 76 people

Tracking posts that say “langfuse”

Track this for your market
Posts, last 24h
26
−16%vs. previous 24h
Posts, last 7 days
157
77 in the last 72h
Velocity
22/day
7-day average
Momentum
+16%
Last 24h vs. a normal day this week

Posts per day, last 30 days

as of Nov 3, 09:00 UTC
100500Oct 20: 31 postsOct 20Oct 21: 52 postsOct 22: 41 postsOct 23: 38 postsOct 24: 0 postsOct 25: 47 postsOct 26: 34 postsOct 26Oct 27: 22 postsOct 28: 29 postsOct 29: 4 postsOct 30: 31 postsOct 31: 23 postsNov 1: 25 postsNov 2: 33 postsNov 2

What changed

as of 27 Oct · checked once a month

First written 27 Oct. When what people say changes, it shows here.

Tone· 61 posts from 53 people· written 27 Oct

How do people feel about Langfuse?

Praise leads reception: 43 of 53 people in the section recommend it, most of them for tracing and evals in one place. 6 value that it is open source and self-hostable, and 4 are still deciding.

Users describe it as the place where tracing, cost, prompts and evals sit together: request traces through tool calls down to the model call, tokens and estimated cost per response, dashboards for volume, latency and errors, and scores from automatic checks or an LLM judge. One team calls it "a game-changer" for improving its AI features; a support engineer says it carried them through a whole debugging week; others describe using trace IDs to close the loop on agent feedback, and reading production traces to find retrieval failures and retry-driven cost.

Self-hosting comes up as its own reason to choose it: a few minutes of `docker compose up`, data kept in the client's own Azure tenant when data residency is required, and "no vendor locked". One caveat sits inside the praise: per-trace cost is handy, but the business side of cost stays your job.

The undecided are few and mostly weighing whether sending everything to their own ClickHouse cluster is enough, without a tool in between.

  • Praises it and recommends it49
  • Values open source and self-hosting6
  • Undecided, planning to try4
  • Satisfied with reservations2
See 5 of 61 posts

a few days ago @brian_armstrong tweeted about adding /feedback endpoints to apis so agents can report bugs and suggest improvements. i think it can become a very big design space for the internet. here is the problem today: >> when an autonomous agent hits an outdated doc, a 404 page or an unclear API, that failure dies in your local terminal. everyone gets zero signal, unless you manually comb through gigabytes of raw session logs. over the last few days, i rolled out a unified feedback architecture across @aeonframework & @miroshark_ entire stack: 6 production apps and 31 agent skills. here is what the /feedback loop actually looks like in production: 1. unified ingestion & enrichment we run a central inbox (node + sqlite on railway). when an agent hits /feedback on any app or skill, the server enriches the report with facts the agent doesn't even know: - exact deployed commit sha - caller session id & langfuse trace id - live route health state - cached run diagnosis if a simulation or background job failed 2. automated credential scrubbing agents paste raw error outputs. error outputs carry authorization headers, jwt tokens, and api keys. we built a multi-tier redactor (vendor key patterns, bearer tokens, private keys) that scrubs every string before it touches disk or logs. 3. semantic deduplication into upvotes if 20 agents hit the same broken route, i dont want 20 separate issues. the server computes word similarity across open reports for the same app. if overlap is >= 60%, it treats the report as an upvote on the existing issue. your backlog automatically ranks itself by how often your agent fleet actually gets blocked by it. 4. closing the loop with worker agents an inbox is useless if a human still has to triage and write the fix by hand. we built an autonomous worker agent in base:0xbf8e8f0e8866a7052f948c16508644347c57aba3 called feedback-builder. it runs in the background, pulls the top-voted open issues, verifies the claim against the actual repository code, writes the fix, and opens a pull request. the full cycle: 1. agent hits an edge case in production 2. agent calls /feedback 3. server attaches commit, trace, and route health 4. fleet upvotes the issue automatically 5. worker agent writes the patch and opens a PR the software factory thesis isn't just letting an llm write code. it's closing the loop so software maintains itself through the agent interface.

Si vous déployez des modèles IA en prod je vous conseille de tester @langfuse pour le monitoring. Il permet de suivre chaque requête de bout en bout. Connaître le coût estimé de chaque réponse et les tokens consommés. Comparer les perfs des modèles et évaluer les réponses avec des scores automatiques ou un LLM juge. Disposer de dashbords avec suivi des volumes, latences, coûts, erreurs. on l'utilise pour l'assistant IA de link et pour l'API du challenge Kiriku. quel outil vous utilisez-vous pour monitorer vos applications IA ?

Thanks, Marc! 🙏 For the record, we use Langfuse in Todoist, and it’s been a game-changer for improving our AI features. We actually have a person in CX (Sergio) who's become a master of the tool 😊 Keep up the good work!

# 模型跑在自己機器上,憑什麼出事了沒人說得清? 模型跑在自己機器上,解決了隱私和成本問題,卻製造了可見性問題:一旦自架模型,沒人告訴你為什麼一個提示跑了 40 秒,也沒人告訴你哪一步檢索餵錯了上下文。解法是 LLM 可觀測性,而最適合本地技術棧的,是自架 Langfuse。這篇教程用 13 步、大約 50 分鐘,從一台空 Ubuntu 機器做到一個被追蹤、被評分、本地託管的 LLM 應用。 ## 一、為什麼需要可觀測性,Langfuse v4 又改了什麼 託管 API 免費送你一個儀表板,本地模型只給你一個終端機和一個希望。一個被追蹤的請求能回答第二週一定會出現的問題:檢索步驟花了多久、提示用了多少 token、模型有沒有撞到輸出上限、上週二的提示修改是讓答案變差了還是只是變得不一樣? McKinsey 2025 年 AI 現況調查發現,62% 的受訪者表示所屬組織至少在實驗 AI Agent。Agent 讓問題成倍放大:一個使用者請求變成五到十次模型呼叫。很多團隊因此同時跑託管模型和本地模型,你需要一個地方記錄每一次呼叫,不管是哪個執行環境提供的。 Langfuse 是開源、可自架的 LLM 應用除錯與迭代平台,最近一季變化很大,四點跟本教程直接相關: - **伺服器 v4 是當前世代**,自架部署已全面可用,Cloud 預定 2026 年 11 月 16 日起只支援 v4,自架使用者自己選遷移日期。 - **SDK 大版本在 2026 年 10 月 14 日一起上線**:Python SDK v5、JS/TS SDK v6,兩者都要求 Cloud 或 v4.59.0 以上版本的自架伺服器。 - **舊 API 刪掉了**:v5、v6 客戶端拿掉了 `api.legacy.*`、`api.ingestion` 這類介面,新增了實驗連結、巢狀元資料篩選、gzip 壓縮的 span 批次。 - **資料模型改為 observation-first**:使用者 ID、會話 ID、元資料、標籤這類關聯屬性,現在存在每個 observation 上,而不只存在 trace 上。 如果你還在跑自架 v3,就不能用新的 SDK 大版本,先留在 Python SDK v4(或 JS/TS SDK v5),直到伺服器升級。本教程假設全新安裝 v4。 ## 二、成品長什麼樣 成品專案叫 TraceLab,五個部分:Docker Compose 跑 Langfuse;本地 OpenAI 相容模型伺服器(Ollama 或 llama.cpp 的 `llama-server`)回答提示;一個小型 Python 套件把每次請求包進 trace,含檢索 span 和生成 span;一個測試套件重放提示檔案,把通過/失敗和延遲分數寫回 Langfuse;一個檢查腳本在你相信任何圖表之前先驗證整條鏈。 每一層記錄的東西、回答的問題: | 層 | 記錄什麼 | 回答什麼問題 | | --- | --- | --- | | Trace | 一次使用者請求、使用者 ID、會話 ID、標籤 | 哪段對話出錯了? | | 檢索 span | 查詢、返回片段、耗時 | 模型看到正確的上下文了嗎? | | 生成 | 提示、完成內容、模型名稱、token 用量 | 花了多少 token 和時間? | | 分數 | trace 上的數字或布林標籤 | 提示版本之間品質是升還是降? | | 會話 | 歸為一次對話的 traces | 多輪對話是怎麼漂移的? | Langfuse 在 2026 年 10 月 13 日重做了會話檢視,可以跨 traces 把使用者和助手訊息顯示為一段連續對話,讓最後一行比一個月前有用得多。 ## 三、先決條件與版本對照 需要一台 Linux 或 macOS 機器,記憶體要夠同時跑 Langfuse 容器和模型。Langfuse 本身是網頁服務、worker、PostgreSQL、ClickHouse、Redis、物件儲存,載入大模型之前先留足餘量。 | 元件 | 用這個版本 | 備註 | | --- | --- | --- | | 作業系統 | Ubuntu 24.04 LTS 或 macOS | Windows 走 WSL 2 | | Docker Engine | 27 以上,含 Compose 插件 | 用 `docker compose version` 檢查 | | Langfuse 伺服器 | v4.59.0 以上 | Python SDK v5 和 JS/TS SDK v6 的硬要求 | | Langfuse Python SDK | v5 系列(當前大版本) | 需要 Python 3.11 以上 | | Python | 3.11 以上 | 3.10 不在 v5 的支援範圍內 | | 本地模型伺服器 | Ollama v0.40.8 以上,或 llama.cpp v0.7.0 | 兩者都提供 OpenAI 相容 API | | Git、OpenSSL | 任何當前版本 | 複製專案和生成密鑰用 | 兩個版本規則值得再看一眼:第一,SDK 和伺服器必須對得上,v4.59.0 之前的 v4 伺服器會拒絕 SDK v5;第二,OpenAI 整合不再支援 1.0 之前的 OpenAI SDK,裝當前的 `openai` 套件。如果現有專案的某個釘選依賴擋住任一升級,就讓那個專案留在 Python SDK v4,給本教程開一個獨立環境。 ## 四、步驟 1 到 3:準備主機、複製 Langfuse、生成密鑰 前面三步故意做得很無聊,因為大多數自架安裝失敗都源於缺工具或密鑰留在預留位置,花十分鐘在這裡,後面省一小時。 **步驟 1:檢查工具。** 跑這五條指令,對照上表,低於最低版本的先升級: ```bash docker --version docker compose version python3 --version openssl version git --version ``` 典型輸出(補丁號會不同):Docker 27.x、Compose v2.x、Python 3.12.x、OpenSSL 3.0.x、git 2.43.x。如果 `python3 --version` 印出 3.10 以下,先裝新版直譯器,Python SDK v5 在 3.10 上裝不上,而且那個報錯很容易被誤讀成網路問題。 **步驟 2:複製 Langfuse 倉庫。** Langfuse 在主倉庫裡發布 Docker Compose 檔案,淺複製就夠,因為你只需要 compose 定義,不需要完整歷史: ```bash mkdir -p ~/tracelab && cd ~/tracelab git clone --depth 1 link cd langfuse docker compose config --services ``` 最後一條指令列出服務名稱,應該看到網頁服務、worker、PostgreSQL、ClickHouse、Redis、S3 相容物件儲存。Langfuse 2026 年 10 月 22 日發布的自架指南建議,給每個部件指定明確負責人,服務清單就是跟團隊展開這場對話的好起點。寫下名字,因為步驟 3 的覆蓋檔案會按名字引用網頁服務,寫作時網頁和 worker 服務分別叫 `langfuse-web` 和 `langfuse-worker`,如果你的輸出不同就相應調整。 **步驟 3:替換每一個預留密鑰。** 隨附的 compose 檔案用 `CHANGEME` 註解標出每個必須改的值,先全部找出來: ```bash grep -n "CHANGEME" docker-compose.yml ``` 用 OpenSSL 生成新值。Langfuse 要求加密金鑰是 64 個十六進位字元,正好是 `openssl rand -hex 32` 印出的長度: ```bash # ENCRYPTION_KEY 用 64 個十六進位字元 openssl rand -hex 32 # NEXTAUTH_SECRET、SALT、資料庫密碼用隨機字串 openssl rand -base64 32 openssl rand -hex 24 ``` 打開 `docker-compose.yml` 替換每個被標記的值。同一個密碼出現兩次的地方(例如 PostgreSQL 服務和網頁、worker 服務裡的連接字串),兩邊必須用相同的值。如果機器能從其他主機連到,公開連接埠要從 `3000:3000` 改成 `127.0.0.1:3000:3000`,因為介面裡會放你的提示和完成內容。 接著在 compose 檔案旁邊建一個 `.env`,放三個應用密鑰,再加一個 `docker-compose.override.yml`,在啟動時建立組織、專案、API 金鑰、第一個使用者,省掉點滑鼠註冊的流程,也讓整套設定以後可以寫成腳本: ```bash cat > .env <<EOF NEXTAUTH_SECRET=<redacted> SALT=$(openssl rand -base64 32) ENCRYPTION_KEY=$(openssl rand -hex 32) LF_PUBLIC_KEY=pk-lf-tracelab-$(openssl rand -hex 8) LF_SECRET_KEY=sk-lf-tracelab-$(openssl rand -hex 16) LF_ADMIN_PASSWORD=<redacted> EOF chmod 600 .env ``` ```yaml # docker-compose.override.yml services: langfuse-web: environment: NEXTAUTH_SECRET: ${NEXTAUTH_SECRET} SALT: ${SALT} ENCRYPTION_KEY: ${ENCRYPTION_KEY} LANGFUSE_INIT_ORG_ID: tracelab-org LANGFUSE_INIT_ORG_NAME: TraceLab LANGFUSE_INIT_PROJECT_ID: tracelab LANGFUSE_INIT_PROJECT_NAME: tracelab-local LANGFUSE_INIT_PROJECT_PUBLIC_KEY: ${LF_PUBLIC_KEY} LANGFUSE_INIT_PROJECT_SECRET_KEY: ${LF_SECRET_KEY} LANGFUSE_INIT_USER_EMAIL: admin@tracelab.local LANGFUSE_INIT_USER_NAME: Admin LANGFUSE_INIT_USER_PASSWORD: ${LF_ADMIN_PASSWORD} langfuse-worker: environment: SALT: ${SALT} ENCRYPTION_KEY: ${ENCRYPTION_KEY} ``` 金鑰對和管理員密碼抄一份放進密碼管理器,步驟 7 的 Python 專案會用到。這個 `.env` 絕對不要提交進 Git。 ## 五、步驟 4 到 5:啟動並驗證健康狀態 **步驟 4:啟動容器。** 第一次啟動會拉取多個映像並跑資料庫遷移,給它幾分鐘: ```bash docker compose up -d docker compose ps docker compose logs -f langfuse-web ``` 看到網頁伺服器就緒的日誌行就可以按 Ctrl+C 離開日誌檢視。健康的 `docker compose ps` 顯示每個服務都在跑、資料庫都是 healthy。網頁容器在 ClickHouse 和 PostgreSQL 跑遷移時重啟一兩次是正常的。如果某個容器持續重啟超過五分鐘,對照疑難排解章節裡對應的症狀處理。 接著確認你實際拉到的伺服器版本,compose 檔案釘住映像標籤,Python SDK v5 需要 v4.59.0 以上: ```bash docker compose images langfuse-web curl -s localhost ``` **步驟 5:驗證健康狀態並登入。** 健康端點會返回一個小 JSON,裡面含伺服器版本: ```json {"status":"OK","version":"4.x.y"} ``` 如果版本以 3 開頭,compose 檔案還指著舊世代。先 `docker compose down` 停掉(不要加 `-v`,這樣資料還在),把網頁和 worker 的映像標籤改到 v4 系列再啟動。GitHub 發布紀錄列出 v4.68.0 在 2026 年 10 月 22 日發布,所以接近這個數字的都算當前版本。 打開 `link `admin@tracelab.local` 和 `.env` 裡的密碼登入,應該落在 `tracelab-local` 專案、空的 traces 表。進設定確認 API 金鑰跟你生成的值一致。如果登入頁提供建立帳號但沒有專案存在,說明初始化變數沒送到網頁容器,用 `docker compose config | grep LANGFUSE_INIT` 查 Compose 實際解析出來的值。 Langfuse 在 2026 年 10 月 7 日加上了 API 金鑰過期功能,預設 30 天、60 天、90 天、一年、自訂日期或不過期。透過初始化變數建立的金鑰跟其他金鑰行為一樣,所以現在就決定要不要輪換:一次性實驗室放著不管沒關係,共用團隊伺服器的話,在介面裡建有過期時間的新金鑰,不要再依賴啟動時那對。 ## 六、步驟 6 到 13:接模型、寫追蹤、跑評分 **步驟 6:跑本地 OpenAI 相容模型伺服器。** Ollama 最新版是 v0.40.9(2026 年 10 月 22 日發布),v0.40 系列為支援的硬體加了自動 MLX 執行;也可以用 llama.cpp 的 `llama-server`。兩個都說 OpenAI 聊天格式,Langfuse 不在乎誰回答,只在乎你的程式碼能不能跟它對話。 **步驟 7:建 Python 專案並接上 Langfuse。** 把 `.env` 裡的公鑰、私鑰貼進專案設定,用 Python SDK v5 的裝飾器把請求包進 trace:一次使用者請求一個 trace,裡面一個檢索 span、一個生成 span。裝飾器寫法讓追蹤程式碼跟業務程式碼分開,之後要關掉追蹤也只要拿掉裝飾器。 **步驟 8 到 10:跑通第一條 trace、掛分數、重放提示檔案。** 先跑一次請求,進介面確認 trace 出現、span 層級正確、token 用量有數字。再給 trace 掛分數:通過/失敗、延遲、品質標籤,分數就是你以後比較提示版本的標尺。最後寫一個測試套件,重放一個提示檔案,把每個提示的通過/失敗和延遲分數寫回 Langfuse,提示一改,整個檔案重跑一遍,品質升降一目了然。 **步驟 11 到 13:檢查腳本、會話檢視、收尾。** 寫一個檢查腳本,在你相信任何圖表之前先驗證整條鏈:伺服器版本對不對、SDK 能不能連上、trace 是不是真的寫進去了。再用 2026 年 10 月 13 日重做的會話檢視,看一次多輪對話是怎麼漂移的。最後把檢查腳本和測試套件一起提交進倉庫,觀測能力才算正式成為專案的一部分,而不是你本機上的一次性實驗。 ## S1哥觀點 自架模型最大的幻覺,是以為「跑起來」就等於「管得住」。跑起來只是第一天,第十天你要回答的問題全是可觀測性問題:哪一步慢了、哪段上下文錯了、哪個提示版本把品質拉下去了。沒有 trace,這些問題的答案全靠猜;有 trace,答案就在表裡。更殘酷的是,SDK 和伺服器版本對不上時,trace 連寫都寫不進去——你以為自己在觀測,其實什麼都沒記下來。問題留給你:你的本地模型今天出問題時,你是打開儀表板,還是打開終端機開始猜?這兩個動作的差距,就是業餘和專業的差距。 原文連結:link 原文:@Tech Insider(由 S1哥 編譯)

Your LLM traces contain your users' data. Prompts, retrieved docs, model outputs: it all lands in your observability tool. Client needs data residency? SaaS tracing breaks that. So we self-host Langfuse in the client's own Azure tenant. Shipping AI isn't just calling an API.

5 of the 61 posts behind this answer, one per person. The other 56, and every new one as it’s posted, come with tracking. Get early access

Pain points· 12 posts from 10 people· written 27 Oct

What do people complain about with Langfuse?

No single complaint leads: the section's five themes are level, each named by 2 people. They run from upgrades that drag and a heavy self-hosted stack to cost figures that come out wrong.

Upgrades take work (2 people): on Python 3.10 the v5 upgrade quietly installs 4.18.0 with no error, and a user who tested it says the guide's ModuleNotFoundError never appeared; prompt migration is described as a task that never ends, though v4 on a self-hosted setup got a positive review.

Running it is heavy going (2 people): several containers to keep up — one comparison puts it at six containers against one process elsewhere — and a dashboard that gets sluggish as features pile up.

Cost numbers go wrong (2 people): cache tokens billed at full price, usage that never lands in the logs, and one report of a cost spike found after the fact, with no alert and no way to tell which customer caused it.

Two people say the spend view answers the wrong question (2): traces show "what you spent, not what you shouldnt have run". Two report defects (2): a write that succeeded while its response was lost, and retries that count waiting time into the queue-time p95.

  • Upgrades and migrations take work4
  • Heavy to run and slow to load2
  • Cost and usage numbers are wrong2
  • Spend view answers the wrong question2
  • Bugs and wrong metrics2
See 5 of 12 posts

« 0 erreur dans Langfuse » - pendant des semaines, j'ai pris ça pour une bonne nouvelle. Ça ne prouve rien. Le déclencheur : un pic de coût découvert après coup, ~440 $ en quatre jours sur une seule feature, sans alerte et sans savoir quel client l'avait causé. 𝗟𝗲 𝘀𝗼𝗰𝗹𝗲 - chaque appel LLM passe par un seul wrapper, qui ouvre une génération Langfuse nommée par feature et écrit une ligne de log applicative, marquée 𝚏𝚊𝚒𝚕𝚎𝚍 si l'appel lève - coût par feature, par modèle et par jour, gratuitement 𝗩𝗲́𝗿𝗶𝗳𝗶𝗲𝗿 𝗹𝗲𝘀 𝗰𝗵𝗶𝗳𝗳𝗿𝗲𝘀 - tokens en cache : Langfuse tarife chaque clé d'usage séparément. Tout remonter sous 𝚒𝚗𝚙𝚞𝚝 comptait le cache plein tarif (-90 % en réalité). Sur la feature du pic, 248 M de tokens d'entrée par semaine affichés à 467 $. Il faut une clé 𝚒𝚗𝚙𝚞𝚝_𝚌𝚊𝚌𝚑𝚎𝚍_𝚝𝚘𝚔𝚎𝚗𝚜 à part. Nos séries d'avant surestimaient d'environ 2× - le tokenizer retombait en silence sur 𝚌𝚕𝟷𝟶𝟶𝚔_𝚋𝚊𝚜𝚎 pour tout modèle inconnu, donc toute la nouvelle génération. Correctif : fixer l'encodage au lieu de le deviner - deux flux sur des modèles sans prix dans Langfuse, affichés à 0 $. Un zéro mérite autant de méfiance qu'un pic 𝗟𝗲 𝗰𝗼𝘂̂𝘁 𝗽𝗮𝗿 𝗰𝗹𝗶𝗲𝗻𝘁 L'id de l'objet métier en metadata de trace. L'API de métriques ne ventile pas par metadata mais filtre dessus : une requête par objet, zéro code. Limite : la trace porte l'objet, pas le compte qui paie. 𝗨𝗻 𝘀𝗲𝘂𝗶𝗹 𝘀𝗶𝗺𝗽𝗹𝗲 3× la moyenne des 7 derniers jours, ou plus de 100 $ dans la journée. Premier déclenchement : 5,5×, à 84 % une seule étude, coût par appel normal. Du volume, une grosse campagne lancée l'après-midi, pas un bug. Le seuil fait regarder quelqu'un, il ne diagnostique pas. 𝗖𝗲 𝗾𝘂𝗲 𝗟𝗮𝗻𝗴𝗳𝘂𝘀𝗲 𝗻𝗲 𝘃𝗼𝗶𝘁 𝗽𝗮𝘀 - pannes OpenAI (échec DNS dans un job, 503 « overloaded » à 2 h) : dans Sentry, pas dans Langfuse - un appel rejeté d'entrée par un nouveau modèle (paramètre de raisonnement refusé) : aucune génération à enregistrer - 2 à 4 % des rapports d'un juge perdus après un changement de modèle : deux JSON collés, erreur avalée, log 𝚌𝚘𝚖𝚙𝚕𝚎𝚝𝚎𝚍 - un tool MCP déclaré mais absent du serveur : OpenAI l'ignore sans erreur. Un redéploiement de staging a retiré tous les tools pendant des heures, rien ne s'est plaint 𝗖𝗲 𝗾𝘂𝗶 𝗿𝗲𝘀𝘁𝗲 - une alerte automatique, aujourd'hui c'est une routine qui lit les métriques chaque matin - le compte client dans les traces - le check des tools manquants ne fait que logger un warning 😅 L'article en réponse 👇

Python 3.10 で Langfuse を使っている方へ。 3.10 では v5 に上がらず、pip はエラーなしで 4.18.0 を入れます。 変わる点を、4.18.0 と 5.0.4 で測りました。 有料部分は、18 項目の例外の型の表です。(有料・100円) link #Langfuse

キュー待ちp95が1〜1.5時間。でも待機中jobは0件。 Langfuse PR #19039は、retry後も残る作成時刻が、前回の実行とbackoffまで「待ち」に数えていたと報告しました。 attemptごとにready・pickup・startを分ける。正しい時刻がなければunknownにする。 AgentOps運用メモ #17 link

5 of the 12 posts behind this answer, one per person. The other 7, and every new one as it’s posted, come with tracking. Get early access

Alternatives· 5 posts· written 27 Oct

What are the alternatives to Langfuse?

No shared rival emerges: the section's five themes are level, each named by 1 person. The comparisons range from LangSmith and Arize Phoenix to self-hosting's upkeep and Claude's built-in tracing.

One person's default is Langfuse itself — "open source, full trace trees for agent runs" — in a three-way split where LangSmith goes deepest for teams already on LangChain and Arize covers monitoring and drift once volume gets serious. Another calls Arize Phoenix the lighter option if you only want traces and evals in one container.

A second line of argument says the tracing is portable: Langfuse, Phoenix and LangSmith all take OpenTelemetry spans over OTLP, so "tracing is a config line" and the real lock-in sits in scores, datasets and prompts that no exporter setting moves.

One person who says they battled data engineering and devops to maintain LLM observability has moved to managed versions only. One notes that Claude Platform's public beta now exports OpenTelemetry traces from the Messages API and SDKs by default, closer to what people used to bolt Langfuse on to get.

  • LangSmith1
  • Arize Phoenix1
  • Tracing is portable, so roll your own1
  • Managed observability instead of self-hosting1
  • Claude's built-in tracing1
See the 5 posts

If you self-host or watch every dollar, Langfuse is my default: open source, full trace trees for agent runs. If you're already on LangChain, LangSmith goes deepest and its traces link into evals. Once volume gets serious, Arize for monitoring and drift.

Langfuse is the closest to LangSmith and self-hosts with docker compose (Postgres, ClickHouse and Redis). Arize Phoenix is lighter if you just want traces and evals in one container. Both take OpenTelemetry, so instrumenting once keeps you from getting locked in.

Langfuse, Phoenix, LangSmith all take OpenTelemetry spans over OTLP. Tracing is a config line. The lock-in is scores, datasets and prompts, because no exporter setting moves them. Pick where your eval loop lives.

why wouldnt something like langfuse pick this up? why would we need a new observability tool here? After battling the entire data engineering and devops to maintain llm observability, I've now moved to only managed versions.

Claude Platform can export OpenTelemetry traces from the Messages API, Managed Agents, and SDKs. Public beta. Default: duration, tokens, model, tool names. Prompts stay opt-in. Closer to what Langfuse users bolt onto GPT agents, now built into Claude. link

Questions· 11 posts from 10 people· written 27 Oct

What do people ask about Langfuse?

Questions about Halo, the agent inside Langfuse, lead with 3 people: where its loop runs, whether state survives a dead sandbox, and whether it sees the same token cost as outside agents. Asking how it differs from other tools and whether traces can carry business context are level behind at 2 people each.

The Halo questions are all about the loop: if it runs outside the sandbox, does every turn spin up a fresh container and lose context, or are tool results checkpointed so a run can replay when a sandbox dies mid-task? They also ask whether the internal loop sees the same token cost and latency as external agents sharing its MCP server, skills and docs.

The comparison questions ask how it differs from Arize AI, Laminar and Braintrust, and pose the blunter "LiteLLM vs LangFuse?". Two ask whether traces can carry what the call itself does not show: what actually landed in an ERP or CRM, intended against actual writes, and who owns a workflow, recorded next to tool calls and permissions.

Singles cover whether a newly shown feature is in the open-source version, whether people reach for Langfuse or their own logging, and how to make its prompt management the source of truth when system prompts and skills live as files in a repo.

  • How does Halo's agent loop work?3
  • How is it different from other tools?2
  • Can traces carry business context?2
  • Langfuse or your own logging?2
  • Is it in the open-source version?1
  • Prompt management as the source of truth1
See 5 of 11 posts

running the agent loop outside the sandbox means state persistence is pure pain, are you using a shared volume or does every turn spin up a fresh container and lose context

5 of the 11 posts behind this answer, one per person. The other 6, and every new one as it’s posted, come with tracking. Get early access

Wishes· 7 posts from 4 people· written 27 Oct

What do people want from Langfuse?

Wishes are spread thin: grading whole runs leads with 2 people, and the other four themes are level behind it. The asks are about evals and metrics tied closer to what actually happened in a run.

On evals, two want runs graded on the sequence of tool calls, not only the final message, since a run can end with the right answer after a write it never should have made. One of them already scores full runs in an external job for wasted tool calls, loops and extra steps, and offers to help with trajectory evals, which they say are on the roadmap.

Two wishes are about linking things back, from one person: an answer should link to the exact traces and filters behind it so someone can inspect why a metric moved, and an expensive turn should link to the tool output that preceded it, which is called more actionable than token totals alone. For evals, one asks to filter by change from baseline as well as an absolute threshold, because a strong item that regresses sharply can still clear the threshold.

Two are about the assistant and prompt workflow: a safety check before it changes any configuration, and weekly prompt changes that don't wait on month-long reviews.

  • Grade whole runs, not just the answer2
  • Link metrics back to the traces2
  • Filter evals by change from baseline1
  • Safety check before the assistant changes things1
  • Ship prompt changes without long reviews1
See 5 of 7 posts

ayy thanks for replying! i saw multi-span / trajectory evals are on the roadmap for @langfuse . ive been doing this via an external job that scores full agent runs (wasted tool calls, loops, extra steps). would love to help out on it, docs, cookbook, testing, code, whatever’s

Linking an answer back to the exact traces and filters would make this especially useful. Then someone can inspect why a metric moved, rather than having to reconstruct the assistant’s query from a summary.

Linking an expensive turn to the tool output that preceded it is much more actionable than token totals alone. Replaying an instruction change on the same tasks would help tell a shorter successful run from one that merely stopped doing useful work.

Filtering by change from baseline would be useful alongside an absolute score threshold. A strong item that regresses sharply can matter even if it remains above the threshold, while a persistently weak item may not be a new regression at all.

5 of the 7 posts behind this answer, one per person. The other 2, and every new one as it’s posted, come with tracking. Get early access

In their words

How do people talk about Langfuse?

The phrases people use most, counted once per person.

  1. 1“mcp server”6
  2. 2“tool calls”4
  3. 3“external agents”4
  4. 4“observability tool”3
  5. 5“cost and latency”2
  6. 6“open source”2
  7. 7“self hosted”2
  8. 8“take opentelemetry”2
  9. 9“assistant inside langfuse”2
  10. 10“agent runs”2
  11. 11“llm observability”2

Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.

Also in LLM observability and evals

Get every post about Langfuse, filtered and labelled

This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.