|
Hi [First name] [First name],
|
|
|
|
October brought skill management to Langfuse: store, version, and trace agent skills the way you already manage prompts. Alongside it, a new session view and API keys you can scope and expire. Here's what's new:
|
|
|
-
Skill management: store and version agent skills next to your prompts.
|
-
Redesigned session view: messages across traces in one continuous conversation.
|
-
Scoped API keys and expiration dates: give each key an Ingestion, Scores ingestion, Viewer, or Full access scope that expires when you choose.
|
-
More decision models: score observations with OpenAI or any System One model.
|
-
Python SDK v5 and JS/TS SDK v6: new majors (version bump for most users).
|
-
Langfuse Government: hardened images and FIPS mode.
|
|
|
|
|
|
🥁 Join us next Thursday, November 12, for the Langfuse Community Town Hall. We'll talk about a big release coming next week and the updated roadmap we're finalizing right now. As always, we'll also demo all the new capabilities we've shipped since the last Town Hall.
|
|
|
|
|
|
|
Skill management
|
|
|
|
|
|
Next to prompts, you can now store and version agent skills in Langfuse. A skill is a SKILL.md with task instructions plus optional supporting files. Create skills in the UI, push a local folder, or publish through the API, CLI, MCP, and SDKs. Each version saves the complete set of files, and labels let you promote or roll back a version without changing application code.
|
|
|
|
With skill tracing configured, generations show which skills have been used. The skill's usage tab combines usage counts, file loads, and linked observations.
|
|
|
|
|
|
|
Redesigned session view
|
|
|
|
|
|
The session view now shows user and assistant messages across traces as one continuous conversation. A sidebar navigates messages, traces, and tool usage, so you can follow where an interaction went off track and jump to the trace behind it.
|
|
|
|
|
|
|
|
|
Scoped API keys and expiration dates
|
|
|
|
|
|
Project API keys now carry permissions: Full access, Viewer, Ingestion, or Scores ingestion. Instrument an application with a key that can send traces but cannot read your prompts or datasets, or collect user feedback from a web app with a key that can only submit scores.
|
|
|
|
You can also now set an expiration date for your keys. Existing keys keep working as they are.
|
|
|
|
|
|
|
More decision models: OpenAI and System One
|
|
|
|
|
|
Decision-model evaluators now run on OpenAI's gpt-6-luna through the Decisions API, and on any model that serves TypeSafe's System One API, such as Subquadratic's SubQ Snap or the open-weights Strands Decider you host yourself.
|
|
|
|
Choice, Score, and Yes/no questions work unchanged, so you can use whichever model fits your cost and latency budget. The Langfuse MCP server also supports decision-model evaluators (API/CLI coming soon).
|
|
|
|
|
|
|
Python SDK v5 and JS/TS SDK v6
|
|
|
|
|
|
Both majors are built for Langfuse v4 and mostly remove APIs that were already deprecated, so for most users the upgrade is a version bump.
|
|
|
|
New: experiment_url / experimentUrl links straight to the comparison view, nested metadata becomes filterable dotted paths, span batches are gzip-compressed, and scores are created in batches. If you self-host, both need Langfuse v4.59.0 or later.
|
|
|
|
|
|
|
Langfuse Government
|
|
For government and other highly regulated environments, Langfuse Self-Hosted Government offers a FIPS mode that runs on FIPS 140-3 validated cryptographic modules and hardened images built on Iron Bank's Red Hat UBI 9.
|
|
|
|
|
|
|
|
|
New Integrations
|
-
Claude Platform: trace the Messages API and Managed Agents with native OpenTelemetry export (docs)
|
-
Cursor: trace Cursor agent sessions, including prompts, tool calls, subagents, and token usage (docs)
|
-
Slack: The Langfuse Slack app is now listed in the Slack Marketplace and lets you get notified for alerts or prompt changes (changelog)
|
|
|
Fixes & Improvements
|
-
Feat: The Experiments page now shows the trade-off between quality, cost, and latency across your runs (changelog)
|
-
Feat: Configure LLM connections once at the organization level and share them with every project; a project connection with the same name takes precedence (changelog)
|
-
Feat: Usage breakdown chart by project or type with CSV export, on the Billing page on Cloud and a new Usage page when self-hosting (changelog)
|
-
Data: Scheduled blob storage exports can include score metadata and provenance, such as the evaluator, annotation queue, and execution trace behind each score (changelog)
|
-
Data: Google Cloud Storage as an export target, with a service account key or ADC (docs)
|
-
Cost: Prompt cache hit rate shown on traces (v4.60.0)
|
-
UX: Cmd+K jumps to prompts, datasets, annotation queues, and evaluators, and offers create actions and recent picks (v4.70.0)
|
-
API: POST /api/public/scores accepts batches (v4.55.0)
|
-
Evals: Evaluator gallery restyled with score previews; evaluator status shows seven-day execution health (v4.60.0)
|
-
Self-hosting: Outbound host allowlists for LLM connections, webhooks, media, and blob storage endpoints (docs)
|
-
Self-hosting: Guide to export application logs to ClickStack (docs)
|
|
|
|
|
Blog
|
-
How we built Halo: Ben on building our in-app assistant on the MCP server, why the agent runs outside its sandbox, and how background runs work
|
-
Customer story: DKB: how DKB keeps an AI banking assistant for around 6 million customers accurate and compliant with self-hosted Langfuse, with engineers and customer service experts working in the same place
|
|
|
|
|
Reminder: Langfuse v4 cutover on November 16
|
|
Langfuse Cloud becomes v4-only on November 16, 2026. Open the Migration Assistant in the sidebar; if it lists required actions for your project, finish them before the cutover.
|
|
|
|
|
|
|
Contribute Full-Time (New: US-East)
|
|
|
|
|
|
Our team is growing (we now hire remote in US-East!). If you're passionate about building open source AI dev tools, we'd love to hear from you.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
-
Developer Relations Engineer, Events & Community (EMEA, AMER)
|
|
|
|
|
|
|
|
|
|
Any feedback, questions, or ideas? Just hit reply; we read and respond to every message.
|
|
|
|
Best,
|
|
Marc
|
|
|
|
—
|
|
|
|
|
|
|
|
|