Langfuse Integration Page

Langfuse integration page creator

This skill scaffolds a new integration page for the langfuse-docs site. Integration pages live as Jupyter notebooks in cookbook/integration_<slug>.ipynb and are converted to MDX by scripts/update_cookbook_docs.sh using the mapping in cookbook/_routes.json. Getting the notebook metadata block, the STEPS_START/STEPS_END wrapper, and the routes entry right is the whole job — once those are correct, the build does the rest.

What to produce

Three things, always, in the user's langfuse-docs checkout:

1. A new notebook at cookbook/integration_<slug>.ipynb that matches the house template (see "Notebook structure" below).
2. A new entry appended to cookbook/_routes.json pointing at the notebook and the target docsPath.
3. A best-effort logo download into public/images/integrations/<slug>_icon.<ext>. If fetching fails, leave a TODO for the user.

Do not run scripts/update_cookbook_docs.sh yourself — that regenerates many files and is slow (~10 min build). The user runs it when they're ready.

Step 1 — Gather what you need, up front

Before writing anything, collect the following. Ask the user for what's missing using a single AskUserQuestion batch where possible. Some answers are mutually exclusive (pick-one); some can be inferred from context.

Always ask (these determine the template and the routes entry):

• Integration name — the human-readable name (e.g., "Pydantic AI", "Fireworks AI", "Temporal"). Used in the title and intro.
• Slug — kebab-case, used in the filename, logo path, and docsPath. Default to the name lowercased with spaces → hyphens, but confirm. Example: "Pydantic AI" → pydantic-ai; "Fireworks AI" → fireworks-ai.
• Category — one of: model-providers, frameworks, gateways, other. This is the <category> segment in docsPath: "integrations/<category>/<slug>". Guidance:
  • model-providers: inference APIs (OpenAI-compatible or otherwise) — Anthropic, Cohere, Fireworks, Groq, Bedrock, Vertex, Gemini, etc.
  • frameworks: agent/app frameworks — LangChain, CrewAI, Pydantic AI, Google ADK, Temporal, Semantic Kernel, etc.
  • gateways: LLM proxies/routers — Portkey, LiteLLM proxy, TrueFoundry, OpenRouter, Kong AI, etc.
  • other: anything else — scraping (Firecrawl, Exa), UIs (Gradio, LibreChat), dev tools, etc.
• Language — python (default) or js. JS integrations use the filename prefix js_integration_<slug>.ipynb and commonly get a -js suffix in the slug when both exist (e.g., anthropic-js, claude-agent-sdk-js).
• Instrumentation pattern — pick one (this determines the template body). See references/patterns.md for full details and match it to the integration:
  • openinference — OpenInference instrumentor library (e.g., openinference-instrumentation-google-adk). Most common for agent frameworks.
  • openai-drop-in — The partner is OpenAI-compatible; use from langfuse.openai import openai. Common for inference providers (Fireworks, Groq, DeepSeek, etc.).
  • framework-native — Framework has built-in instrumentation hook (e.g., Agent.instrument_all() for Pydantic AI).
  • otel-direct — Partner emits OTel natively; configure an OTLP exporter pointing at Langfuse. Less common; used for things like Temporal, some MLflow setups.

Ask if not obvious:

• Intro blurb about the partner — one sentence ("What is X?"). If the user didn't give one and there's a URL, you can draft it and confirm.
• Logo source — if the user provided a URL, great; if not, see Step 4 for the fetch heuristic.
• Install command, env vars beyond the Langfuse ones, and a minimal runnable example — needed for the code cells. If missing, draft from docs and mark as TODO: confirm.

How to ask

Use AskUserQuestion with options formatted as the four categories and four patterns. Keep the number of questions ≤ 4. If the user gave full context (e.g., they pasted a complete code example and mentioned the framework), skip questions you can answer confidently from context and just confirm in your response.

Step 2 — Generate the notebook

You have two ways to create the .ipynb:

1. Preferred: use the bundled builder script scripts/build_notebook.py. It takes a structured JSON/YAML description of the cells and writes a properly formatted notebook. Using the script avoids subtle JSON errors (trailing commas, missing "source" arrays, line-split source strings) that break nbconvert.

   python3 <skill-dir>/scripts/build_notebook.py \
     --out cookbook/integration_<slug>.ipynb \
     --spec /tmp/<slug>_spec.json

   See the script's --help for the spec schema. There are examples at the bottom of references/patterns.md.

2. Fallback: write the .ipynb file directly. If you do this, open an existing notebook (e.g., cookbook/integration_pydantic_ai.ipynb) first and mirror its JSON shape exactly. Be careful: every source field is a list of strings, each ending in \n except the last; markdown cells carry "metadata": {"vscode": {"languageId": "raw"}}; code cells carry "execution_count": null, "outputs": [].

Whichever route you pick, the cell structure must match the house template.

Notebook structure (the template)

Every integration page has the same skeleton. Section order matters because the MDX converter in scripts/move_docs.py reads the NOTEBOOK_METADATA comment from the top of the first cell and wraps everything between STEPS_START and STEPS_END in a <Steps> component.

Cell 1 — markdown. Metadata + intro.

The first line is a single-line HTML comment with all the page metadata. Attribute format is key: "value" (double-quoted), space-separated, on one line. Required keys:

<!-- NOTEBOOK_METADATA source: "⚠️ Jupyter Notebook" title: "<Page title>" sidebarTitle: "<Short nav label>" logo: "/images/integrations/<slug>_icon.<ext>" description: "<1-sentence SEO description>" category: "Integrations" -->

Then the page H1, a 1-sentence intro, and two blockquote callouts:

# Integrate Langfuse with <Partner Name>

This notebook shows how to integrate **Langfuse** with **<Partner>** to [monitor / debug / trace / evaluate] your LLM application.

> **What is <Partner>?** [<Partner>](<partner url>) is <one sentence about the partner>.

> **What is Langfuse?** [Langfuse](https://langfuse.com) is an open-source LLM engineering platform that helps teams trace, debug, and evaluate their LLM applications.

Title-writing notes: prefer "Observability for <Partner> with Langfuse" for model providers and inference APIs, "Integrate Langfuse with <Partner>" for frameworks, and "Trace <Partner> Workflows with Langfuse" for orchestration tools. Sidebar title is the short name (e.g., "Pydantic AI", "Fireworks AI", "Temporal").

Cell 2 — markdown. Start of steps.

<!-- STEPS_START -->
## Step 1: Install Dependencies

Cell 3 — code. Install.

%pip install langfuse <partner-package> -U

Use -U to upgrade. For JS notebooks, use npm install in a shell cell (see the JS examples in cookbook/js_integration_*.ipynb).

Cell 4 — markdown. Env var setup prose.

One short paragraph mentioning that keys come from Langfuse project settings, linking to Langfuse Cloud and https://langfuse.com/self-hosting.

Cell 5 — code. Env vars.

Always include the three Langfuse vars in this exact shape (EU active by default, other regions noted in a comment) plus whatever the partner needs. Every os.environ.setdefault line must end with ; — setdefault returns the live env value (a real key, if one is already set), and the semicolon keeps Jupyter from echoing it into the saved cell output:

import os

# Get keys for your project from the project settings page: https://langfuse.com/cloud
os.environ.setdefault("LANGFUSE_PUBLIC_KEY", "pk-lf-...");
os.environ.setdefault("LANGFUSE_SECRET_KEY", "sk-lf-...");
os.environ.setdefault("LANGFUSE_BASE_URL", "https://cloud.langfuse.com"); # 🇪🇺 EU region (API host)
# Other Langfuse data regions include 🇺🇸 US: https://us.cloud.langfuse.com, 🇯🇵 Japan: https://jp.cloud.langfuse.com and ⚕️ HIPAA: https://hipaa.cloud.langfuse.com

# <Partner> API key
os.environ.setdefault("<PARTNER>_API_KEY", "...");

Cell 6 — markdown + cell 7 — code. Initialize Langfuse client with auth check. (Skip this pair for the openai-drop-in pattern, which relies on the langfuse OpenAI wrapper instead.)

from langfuse import get_client

langfuse = get_client()

# Verify connection
if langfuse.auth_check():
    print("Langfuse client is authenticated and ready!")
else:
    print("Authentication failed. Please check your credentials and host.")

Cells 8+ — Instrumentation + runnable example. These are pattern-specific. See references/patterns.md for the exact cell bodies for each of the four patterns.

Final steps cell — markdown. View traces.

## Step N: View Traces in Langfuse

After running the example, open [Langfuse Cloud](https://langfuse.com/cloud) to see the trace, including prompts, completions, tool calls, token usage, and latency.

![Example trace in Langfuse](https://langfuse.com/images/cookbook/integration-<slug>/<slug>-example-trace.png)

<!-- TODO: replace with your actual trace screenshot (upload to langfuse.com images) and example trace link -->
[Example trace in Langfuse](<example trace URL or placeholder>)

<!-- STEPS_END -->

Last cell — markdown. LearnMore.

<!-- MARKDOWN_COMPONENT name: "LearnMore" path: "@/components-mdx/integration-learn-more.mdx" -->

For JS integrations use @/components-mdx/integration-learn-more-js.mdx instead.