---
title: Mixpanel for LLM Apps with Langfuse
sidebarTitle: Mixpanel
logo: /images/integrations/mixpanel_icon.svg
description: Display your Langfuse metrics in Mixpanel dashboards.
---
# Mixpanel Integration
[Mixpanel](https://mixpanel.com) is a popular choice for product analytics. While Langfuse offers [metrics](/docs/metrics/overview) out of the box, many of our users have asked for a way to **integrate their LLM related metrics that they capture with Langfuse into their Mixpanel dashboards**.
End-to-end demo of the integration together with Ryan from Mixpanel:
We've built an integration to make it easy to answer questions like:
- _"Are my most active users also the ones who are most engaged with my LLM content?"_
- _"Does interacting with the LLM feature relate to higher retention rates?"_
- _"How does the LLM feature impact my conversion rates?"_
- _"Does the user feedback that I capture in Langfuse correlate with the user behavior that I see in Mixpanel?"_
## Example dashboard
We've enabled the integration for the [Langfuse example project](/demo) and created a dashboard using the `AI Company KPIs` template by Mixpanel. Dashboard preview ([public link](https://eu.mixpanel.com/p/V92wFA5YwNjKn9K9vcUmAX)):
## Get started
### Enable the integration
Configure this integration in your Langfuse project settings. You will need to select your Mixpanel region and provide your Mixpanel Project Token.
**Available Regions:**
- US (api.mixpanel.com)
- EU (api-eu.mixpanel.com)
- India (api-in.mixpanel.com)
You can find your Project Token in your Mixpanel project settings under "Project Settings" > "Project Token".
### Initial sync
Once integrated, Langfuse will sync all historical data from your project to Mixpanel. After the initial sync, new data is automatically synced every hour (with a 30-minute delay) to keep your Mixpanel dashboards up to date.
### Build a dashboard in Mixpanel
Once integrated, you can build dashboards in Mixpanel to visualize your Langfuse metrics (see [reference below](#details)).
You can use the `AI Company KPIs` template to get started with a pre-built dashboard that includes many common metrics. See the dashboard preview above for an example.
## Integration details [#details]
On an hourly schedule, Langfuse batches aggregated events and metrics to your Mixpanel instance.
### Metadata matching
Matching of metadata helps to join the data from Langfuse with the data from Mixpanel:
| Langfuse | Mixpanel | Notes |
| ----------------------------------------------- | ------------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| [`user_id`](/docs/observability/features/users) | `distinct_id`, `$user_id` | |
| `observation.start_time`, `score.timestamp` | `time` | Sent as milliseconds since epoch. For [legacy events](#export-source): `trace.timestamp` and `generation.started_at`. |
| `trace.metadata.$mixpanel_session_id` | `session_id` | Optional, you need to add the Mixpanel session id to your Langfuse trace [metadata](/docs/observability/features/metadata). |
### Events [#events]
The integration sends one `[Langfuse] Observation` event per observation and one `[Langfuse] Score` event per score.
Is there any additional information that would be helpful? You can request more events or properties [here](/ideas).
Projects using a legacy export source receive different events — see [Legacy
export sources](#export-source) below.
#### Event: `[Langfuse] Observation` [#event-observation]
One event per observation, covering every observation type.
| Property | Description |
| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------- |
| `time` | Milliseconds since epoch when the observation started. |
| `distinct_id` | User ID or anonymous identifier. |
| `$user_id` | User ID sent to Mixpanel's native user ID field. |
| `$insert_id` | Unique identifier for deduplication. |
| `session_id` | Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id). |
| `langfuse_observation_name` | The name of the observation. |
| `langfuse_type` | The observation type: `GENERATION`, `SPAN`, `EVENT`, `AGENT`, `TOOL`, `CHAIN`, `RETRIEVER`, `EVALUATOR`, `EMBEDDING`, or `GUARDRAIL`. |
| `langfuse_trace_name` | Name of the trace related to the observation. |
| `langfuse_trace_id` | The unique identifier of the trace related to the observation. |
| `langfuse_url` | The URL of the observation on the host platform. |
| `langfuse_user_url` | Deeplink to the user profile in Langfuse. |
| `langfuse_id` | Unique identifier of the observation. |
| `langfuse_cost_usd` | Computed total cost of the observation. |
| `langfuse_input_units` | Number of tokens utilized in prompting (for LLM calls). |
| `langfuse_output_units` | Number of tokens produced (for LLM calls). |
| `langfuse_total_units` | Total number of tokens consumed (for LLM calls). |
| `langfuse_session_id` | The session ID associated with the trace of the observation. |
| `langfuse_project_id` | The ID of the project where the observation happened. |
| `langfuse_project_name` | The name of the project where the observation happened. |
| `langfuse_user_id` | The user ID that started the trace linked to the observation. In case it's unavailable, it defaults to `langfuse_unknown_user`. |
| `langfuse_latency` | The observed latency of the observation. |
| `langfuse_time_to_first_token` | The time taken to generate the first token when streaming. |
| `langfuse_release` | Release information of the trace attached with the observation. |
| `langfuse_version` | The versioning info about the observation. |
| `langfuse_model` | The model used (for LLM calls). |
| `langfuse_level` | The level associated with the observation. |
| `langfuse_tags` | Any tags attached to the trace of the observation. |
| `langfuse_environment` | The environment associated with the observation. |
| `langfuse_event_version` | The integration version with Langfuse. |
#### Event: `[Langfuse] Score` [#event-score]
One event per score.
| Property | Description |
| ----------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| `time` | Milliseconds since epoch when the score event occurred. |
| `distinct_id` | User ID or anonymous identifier. |
| `$user_id` | User ID sent to Mixpanel's native user ID field. |
| `$insert_id` | Unique identifier for deduplication. |
| `session_id` | Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id). |
| `langfuse_score_name` | The name associated with the score. |
| `langfuse_score_value` | The value of the score. |
| `langfuse_score_string_value` | The string value of the score. For BOOLEAN and CATEGORICAL scores, this will be the string representation of the value. |
| `langfuse_score_data_type` | The data type of the score (NUMERIC, BOOLEAN, CATEGORICAL). |
| `langfuse_score_comment` | Any comments attached with the score. |
| `langfuse_score_metadata` | Any metadata attached with the score. |
| `langfuse_score_entity_type` | The entity the score is attached to: `trace`, `session`, `dataset_run`, or `unknown`. |
| `langfuse_dataset_run_id` | The unique identifier of the dataset run, for scores attached to a dataset run. |
| `langfuse_trace_name` | The name of the trace associated with the score. |
| `langfuse_trace_id` | The unique identifier of the trace associated with the score. |
| `langfuse_user_url` | Deeplink to the user profile in Langfuse. |
| `langfuse_id` | The unique identification of the score. |
| `langfuse_session_id` | The session identification related to the score's trace. |
| `langfuse_project_id` | The project identification linked with the score's trace. |
| `langfuse_project_name` | The name of the project linked with the score's trace. |
| `langfuse_user_id` | The user ID that triggered the trace tied with the score. If not available, defaults to `langfuse_unknown_user`. |
| `langfuse_release` | The release information of the trace associated with the score. |
| `langfuse_tags` | Any tags related to the trace of the score. |
| `langfuse_environment` | The environment associated with the score. |
| `langfuse_event_version` | The integration version with Langfuse. |
## Legacy export sources [#export-source]
Cloud projects created before 2026-05-20 and self-hosted deployments have an `Export Source` selector in the integration settings (introduced with the Langfuse v4 preview). Cloud projects created on or after 2026-05-20 always use `Enriched observations (recommended)` — the schema documented [above](#events) — and can skip this section.
Available options:
- `Enriched observations (recommended)`: sends the [`[Langfuse] Observation`](#event-observation) events documented above. Provides significantly better export performance and covers every observation type.
- `Traces and observations (legacy)`: sends the [`[Langfuse] Trace`](#event-trace) and [`[Langfuse] Generation`](#event-generation) events documented below.
- `Traces and observations (legacy) and enriched observations`: sends all of the above. This creates duplicate records by design, as a validation aid during migration.
[`[Langfuse] Score`](#event-score) events are always sent, regardless of the selected export source.
`Traces and observations (legacy)` sources may be deprecated in the future. All new export jobs should use `Enriched observations (recommended)`, and existing legacy jobs are strongly recommended to upgrade.
### Migrating from the legacy export source [#migrate-export-source]
New integrations default to `Enriched observations (recommended)`; existing integrations continue to use `Traces and observations (legacy)` until changed.
November 16, 2026 (2026-11-16) is the Langfuse Cloud v4 cutover. If you have not switched
by then, remaining legacy exports switch to `Enriched observations`
automatically. Migrate on your own schedule beforehand to avoid an
unannounced hard cliff in your event stream. If you are self-hosting, the
switch happens automatically once you run the v4 server in `events_only`
mode.
Mixpanel is a schemaless event store, so no setup is needed on the Mixpanel side — the new event name and properties appear automatically once you switch. All migration effort goes into the assets built on top of the events: reports, boards, funnels, cohorts, alerts, and warehouse syncs.
Keep in mind:
1. **Nothing is backfilled.** The sync is watermark-based: the old event streams stop at the switch date and the new one starts. Reports keyed on the old event names show a hard cliff — add a Mixpanel annotation at the cutover date and keep old boards around for lookback.
2. **Queries need rewriting, not re-pointing.** Reports on `[Langfuse] Generation` become reports on `[Langfuse] Observation` filtered by `langfuse_type = 'GENERATION'`. Per-trace metrics (total cost, observation count, latency) must be recomputed by aggregating observation events over `langfuse_trace_id`, since there is no trace-summary event anymore — see [changes from legacy to enriched events](#legacy-vs-enriched).
3. **The combined mode double-counts by design.** With `Traces and observations (legacy) and enriched observations`, a generation is exported both as a `[Langfuse] Generation` and as a `[Langfuse] Observation` event. Mixpanel's `$insert_id` deduplication is scoped per event name, so these are not collapsed, and metrics that sum cost or tokens across both event types double-count during migration. Use this mode only to build and validate new boards side by side.
4. **The switch is effectively one-way.** Once legacy writes are off, the legacy sources can no longer be selected.
To migrate:
1. Switch to `Traces and observations (legacy) and enriched observations`.
2. Duplicate your reports and boards onto the `[Langfuse] Observation` event: filter by `langfuse_type`, use `langfuse_observation_name`, and rebuild per-trace metrics as aggregations over `langfuse_trace_id`.
3. Validate the new numbers against the old ones side by side. Expect intentional coverage differences: the enriched source includes all observation types, not just generations.
4. Switch to `Enriched observations (recommended)` and annotate the cutover date on any legacy boards you keep.
For rollout details, see the [enriched observations default changelog](/changelog/2026-05-20-blob-storage-enriched-default).
### Changes from legacy to enriched events [#legacy-vs-enriched]
- **`[Langfuse] Generation` is replaced by `[Langfuse] Observation`.** The two events are field-for-field identical, with two exceptions: `langfuse_generation_name` is renamed to `langfuse_observation_name`, and a new `langfuse_type` property carries the observation type.
- **Coverage expands to all observation types.** The legacy source only exports generations; the enriched source exports every observation type. Expect a higher event volume after switching — Mixpanel bills per event. Filter on `langfuse_type` = `GENERATION` to reproduce the legacy scope.
- **No standalone trace event.** Trace context is available on every observation event via `langfuse_trace_id` and `langfuse_trace_name`. The trace-level rollup properties have no enriched equivalent: `langfuse_count_observations`, the trace-level `langfuse_cost_usd` and `langfuse_latency` aggregates, and the trace-scoped `langfuse_url`. To approximate them, aggregate observation events by `langfuse_trace_id`.
- **Score events are identical** in both sources.
- **Identity and deduplication are unchanged**: `distinct_id` is still derived from `langfuse_user_id`, `session_id` and `$insert_id` handling is the same, and `langfuse_event_version` remains `1.0.0` for all events.
### Event: `[Langfuse] Trace` [#event-trace]
Sent only by the legacy export sources. There is no enriched equivalent — see [changes from legacy to enriched events](#legacy-vs-enriched).
| Property | Description |
| ----------------------------- | ------------------------------------------------------------------------------------------------------------ |
| `time` | Milliseconds since epoch when the event occurred. |
| `distinct_id` | User ID or anonymous identifier. |
| `$user_id` | User ID sent to Mixpanel's native user ID field. |
| `$insert_id` | Unique identifier for deduplication. |
| `session_id` | Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id). |
| `langfuse_trace_name` | The name of the trace. |
| `langfuse_url` | The URL of the trace. |
| `langfuse_user_url` | Deeplink to the user profile in Langfuse. |
| `langfuse_id` | The unique identification of the trace. |
| `langfuse_cost_usd` | The total cost associated with the trace. |
| `langfuse_count_observations` | The number of observations in the trace. |
| `langfuse_session_id` | The session ID related to the event. |
| `langfuse_project_id` | An ID associated with the project related to the event. |
| `langfuse_project_name` | The name of the project related to the event. |
| `langfuse_user_id` | User ID related to the event. If it's null, it's marked as langfuse_unknown_user. |
| `langfuse_latency` | The latency of the event. |
| `langfuse_release` | Information relating to the event release. |
| `langfuse_version` | The version of the event. |
| `langfuse_tags` | Any tags associated with the event. |
| `langfuse_environment` | The environment associated with the trace. |
| `langfuse_event_version` | The integration version of Langfuse. |
### Event: `[Langfuse] Generation` [#event-generation]
Sent only by the legacy export sources. Only covers observations of type generation; replaced by [`[Langfuse] Observation`](#event-observation) in the enriched source.
| Property | Description |
| ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------ |
| `time` | Milliseconds since epoch when the generation started. |
| `distinct_id` | User ID or anonymous identifier. |
| `$user_id` | User ID sent to Mixpanel's native user ID field. |
| `$insert_id` | Unique identifier for deduplication. |
| `session_id` | Optional session identifier (from `$mixpanel_session_id` in metadata, or falls back to Langfuse session_id). |
| `langfuse_generation_name` | The name of the generation. |
| `langfuse_trace_name` | Name of the trace related to the generation. |
| `langfuse_trace_id` | The unique identifier of the trace related to the generation. |
| `langfuse_url` | The URL of the generation on the host platform. |
| `langfuse_user_url` | Deeplink to the user profile in Langfuse. |
| `langfuse_id` | Unique identifier of the generation. |
| `langfuse_cost_usd` | Computed total cost of the generation. |
| `langfuse_input_units` | Number of tokens utilized in prompting the generation. |
| `langfuse_output_units` | Number of tokens produced by the generation. |
| `langfuse_total_units` | Total number of tokens consumed in the generation process. |
| `langfuse_session_id` | The session ID associated with the trace of the generation. |
| `langfuse_project_id` | The project identification where the generation happened. |
| `langfuse_project_name` | The name of the project where the generation happened. |
| `langfuse_user_id` | The user ID that started the trace linked to the generation. In case it's unavailable, it defaults to `langfuse_unknown_user`. |
| `langfuse_latency` | The observed latency of the generation. |
| `langfuse_time_to_first_token` | The time taken to generate the first token when streaming. |
| `langfuse_release` | Release information of the trace attached with the generation. |
| `langfuse_version` | The versioning info about the generation. |
| `langfuse_model` | The model used during this generation's process. |
| `langfuse_level` | The level associated with the generation. |
| `langfuse_tags` | Any tags attached to the trace of the generation. |
| `langfuse_environment` | The environment associated with the generation. |
| `langfuse_event_version` | The integration version with Langfuse. |
## Troubleshooting
**Missing data in Mixpanel?** Please check that you have correctly selected the Mixpanel region and entered the Project Token in your Langfuse project settings. The integration syncs data every hour with a 30-minute delay, so there may be a short lag before new events appear. Reach out to us if you encounter any other issues with the integration.
---
## Agent Instructions
This page is part of the [Langfuse](https://langfuse.com) documentation, published as plain Markdown for AI agents. Every page is available as Markdown by appending `.md` to its URL, or by sending an `Accept: text/markdown` header. This page: `https://langfuse.com/integrations/analytics/mixpanel.md`.
### Querying these docs
If the answer is not on this page, query the documentation instead of guessing:
- **Semantic search** across all Langfuse docs, returning an answer with the relevant pages and excerpts. Ask a specific, self-contained question:
```bash
curl -sG "https://langfuse.com/api/search-docs" --data-urlencode "query=How do I trace a LangGraph agent?"
```
- **Index of every page**: , with per-section indexes [llms-docs.txt](https://langfuse.com/llms-docs.txt), [llms-integrations.txt](https://langfuse.com/llms-integrations.txt), and [llms-self-hosting.txt](https://langfuse.com/llms-self-hosting.txt).
### Before writing Langfuse code
- **Install the [Langfuse Agent Skill](https://langfuse.com/docs/api-and-data-platform/features/agent-skill).** It encodes Langfuse's own best practices for instrumentation, prompt management, and evaluation, and materially improves results.
- **Read [What does a good trace look like?](https://langfuse.com/docs/observability/best-practices.md)** before instrumenting an application.
- **Verify endpoints, parameters, and response fields** against the [API reference](https://api.reference.langfuse.com) instead of inferring them from code examples.
- **Use the [Langfuse CLI](https://langfuse.com/docs/api-and-data-platform/features/cli)** (`npx langfuse-cli api `) to read or write traces, prompts, datasets, and scores from the terminal.
Found an error in these docs? Please open an issue at .