We Open-Sourced Our Langfuse Skills — Operate LLM Observability from Your Coding Agent

We previously shared how we built a debugging environment using Claude Code × Langfuse. We've now open-sourced those internal Langfuse integration skills as langfuse-skills.

Why We Open-Sourced It

After publishing the previous article, we received many requests from developers who wanted to use the same setup in their own projects.

If you develop LLM applications, you know the drill: constantly switching between your Langfuse dashboard and your editor. Opening a browser to check a trace, copying IDs, digging through nested JSON. The idea of letting your coding agent handle this resonated with a lot of developers.

So we packaged our internal skills into a reusable package compatible with the Vercel skills ecosystem.

Installation

One line is all it takes.

# Claude Code
npx skills add neuradex/langfuse-skills -a claude-code

# Cursor
npx skills add neuradex/langfuse-skills -a cursor

# Codex / Windsurf / Cline / Roo Code and more
npx skills add neuradex/langfuse-skills -a <agent>

Set your environment variables and you're good to go.

export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."

What You Can Do

We provide 10 skills.

Trace & Session Analysis: langfuse-trace-list / langfuse-trace-view / langfuse-session-list / langfuse-session-view / langfuse-observation-view — Diagnose slow traces, analyze conversation quality, and monitor costs using natural language.

Prompt Management: langfuse-prompt-list / langfuse-prompt-view / langfuse-prompt-upsert — Review, improve, and deploy prompts entirely through conversation with your agent. When you identify a prompt as the root cause, fix it on the spot and save with a development label.

Model & Cost Management: langfuse-model-list / langfuse-model-upsert — View and update model pricing configurations.

How It Works in Practice

The workflow from our previous article works out of the box.

You: "This trace is slow. Find the cause."
Agent: (fetches trace via langfuse-trace-view)
       "Input tokens exceed 12,000.
        The system prompt has too many few-shot examples.
        I suggest reducing from 3 to 1."
You: "Fix the prompt."
Agent: (fetches current prompt via langfuse-prompt-view)
       (proposes improvements)
You: "OK, deploy it."
Agent: (saves via langfuse-prompt-upsert)
       "Saved with the development label."

From trace analysis to prompt fixes, all without leaving your editor.

Wrapping Up

langfuse-skills is a skill package that integrates LLM observability into your coding agent's workflow.

GitHub: neuradex/langfuse-skills
Supported Agents: Claude Code, Cursor, Codex, Windsurf, Cline, Roo Code, GitHub Copilot
License: MIT

If you're tired of switching between your dashboard and editor while developing LLM applications, give it a try.