Mixpanel Integration

Mixpanel is a popular choice for product analytics. While Langfuse offers metrics 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 and created a dashboard using the AI Company KPIs template by Mixpanel. Dashboard preview (public link):

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).

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

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`	`distinct_id`, `$user_id`
`observation.start_time`, `score.timestamp`	`time`	Sent as milliseconds since epoch. For legacy events: `trace.timestamp` and `generation.started_at`.
`trace.metadata.$mixpanel_session_id`	`session_id`	Optional, you need to add the Mixpanel session id to your Langfuse trace metadata.

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.

Projects using a legacy export source receive different events — see Legacy export sources below.

Event: `[Langfuse] 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`

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

Cloud projects created before 2026-05-20 and self-hosted deployments have an Export Source selector in the integration settings (introduced with Langfuse v4, Fast Preview). Cloud projects created on or after 2026-05-20 always use Enriched observations (recommended) — the schema documented above — and can skip this section.

Available options:

Enriched observations (recommended): sends the [Langfuse] Observation events documented above. Provides significantly better export performance and covers every observation type.
Traces and observations (legacy): sends the [Langfuse] Trace and [Langfuse] 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 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

New integrations default to Enriched observations (recommended); existing integrations continue to use Traces and observations (legacy) until changed.

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:

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.
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.
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.
The switch is effectively one-way. Once legacy writes are off, the legacy sources can no longer be selected.

To migrate:

Switch to Traces and observations (legacy) and enriched observations.
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.
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.
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.

Changes from legacy to enriched events

[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`

Sent only by the legacy export sources. There is no enriched equivalent — see changes from legacy to enriched events.

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`

Sent only by the legacy export sources. Only covers observations of type generation; replaced by [Langfuse] 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.

Was this page helpful?

On this page