Langfuse 4.69: Skill Tracing, MCP Updates and AI Gateway Setup

Langfuse released version 4.69.0 on October 23 with clearer agent skill tracing, MCP tools for managing skill versions and a guided AI Gateway setup. For developers debugging agents or running Langfuse themselves, the useful changes are in what can be inspected and what must be configured.

The release builds on existing Skills and AI Gateway work. Both were already present in earlier versions. This is a workflow update, rather than the first launch of either product, and the release notes do not announce a pricing change or guarantee access for every Cloud organization.

Skill tracing separates availability from requests

An agent having access to a skill does not mean it asked to use it. Langfuse's generation preview now makes that distinction visible instead of presenting one undifferentiated list of available skills.

The tracing implementation splits skills into Requested and Not requested based on the generation's output tool calls. Expanding a requested skill reveals the distinct file paths it asked for. A direct skill request shows SKILL.md; supported resource and file-read calls retain their requested paths.

Requested does not mean successfully loaded. Existing loaded-skill information remains separate. The matching also has limits: malformed arguments, ambiguous version matches and historical calls in the input are excluded. Generic file reads need a recognizable skills/<name>/ resource path; the interface cannot infer arbitrary custom directories.

That distinction matters when a support agent gives the wrong answer. The next question is whether it ignored an available policy, requested the wrong file or failed while loading the right one. Those are different problems, even if the final answer looks equally wrong.

Finding traces that used a particular skill

Version 4.69 also adds separate available-skill and loaded-skill name filters to the trace sidebar. The Users table and a user's trace table inherit those filters, while Sessions are outside this change.

The search fields are availableSkills and loadedSkills, synchronized with the sidebar. This gives teams a way to narrow the evidence before reading individual traces, rather than assuming that every run with a skill available actually used it.

When inspecting child skill observations, check the Is Root Observation filter. Langfuse's reproduction instructions clear it to include child spans. Otherwise, an apparently missing match may be a view restriction rather than evidence that the agent never loaded the skill.

MCP can create and promote skill versions

The MCP update adds createSkillVersion, updateSkillLabels and updateSkillTags. On configured instances with Halo and Skills enabled, Halo can use these operations instead of only reading skills.

Creation supplies a complete immutable file set. Labels can be changed while preserving the automatic latest label and protected-label permissions. Tag updates replace tags across all versions without changing their labels, and discovery tags can affect which skills an integration loads.

These are writes with operational consequences, so Halo requires approval for all three tools. Promoting a version to production also requires an explicit user request. That is an important boundary: asking an assistant to improve a skill should not silently make the revised instructions the production version.

AI Gateway setup becomes more explicit

The guided setup change brings gateway configuration and provider credentials into one Setup page. Unconfigured organizations move through recording preferences, a provider credential and optional team invitations. Recording choices remain None, Usage and Full.

Separately, the key-creation dialog adds instructions for API use, Claude Desktop, Claude Code, Codex and OpenCode. The setup snippets reflect enabled providers and use a key placeholder, rather than embedding the newly created secret in the instructions.

For self-hosted operators, the most important change is below the interface. The versioned deployment configuration introduces LANGFUSE_AI_GATEWAY_DEPLOYED, defaulting to false. Set it to true only once the gateway is actually deployed alongside Web. Enabling gateway access alone does not override that deployment check.

LANGFUSE_AI_GATEWAY_PUBLIC_URL can specify the public gateway origin shown to users. It accepts an HTTP or HTTPS origin, without credentials, a path, query or fragment. It is not a place to append an API route such as /v1. These settings describe deployment and client configuration; changing them does not deploy a gateway for you.

What teams should check after upgrading

For agent teams, inspect one failed run across available, requested and loaded skills before rewriting the instructions. For self-hosted teams, verify gateway deployment state and the public origin before distributing client setup steps.

The useful direction here is more precise diagnosis. A skill that was offered, a file that was requested and instructions that actually reached the agent are different pieces of evidence. Keeping them separate makes it easier to change the right part of the system.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI and developer-tool teams turn technical updates and product launches into clear stories that developers and founders read and share.

Book a 30-minute call.