Recipe: try Kubit on your Pydantic Logfire traces

This recipe runs a proof of concept on traces you already have in Pydantic Logfire. You export them to a file with Kubit’s open-source trace-export tool, then upload the file to a Sandbox workspace. Your app, its instrumentation and Logfire stay as they are. It takes about 30 minutes. A Sandbox workspace takes uploaded trace files only. It has no ingestion pipeline and no API key. To send live traces later, follow the Quickstart in a new Agent workspace.

Before you start

A Logfire project with traces from your agent. Pydantic AI traces work best, because Pydantic AI already writes the OpenTelemetry GenAI attributes Kubit reads.
Permission to create a read token in that Logfire project
Python 3.10 or later, and git
A Kubit account that can create workspaces. Sign-up needs an early-access link from Kubit or an invite from your organization. See Set up your organization.

Prompt for your coding agent

Paste this into Claude Code, Cursor or Codex. It runs the export and leaves the token and the upload to you:

Read https://learn.kubit.ai/agent-observability/cookbook/logfire-sandbox.md.
Clone trace-export and install the Logfire exporter, as in the steps "Install trace-export" and "Export the traces".
I'll put LOGFIRE_READ_TOKEN in pydantic-logfire/.env myself. Don't print the token or ask me for it.
Export spans from <since> to <until> into data/logfire.otel.jsonl and show me the summary.

Steps

Create a Logfire read token

In Logfire, open the project with your traces, then go to its settings and Read tokens. Create a token and copy it.A read token can read one project only. trace-export works out the region (US or EU) from the token.

Install trace-export

Clone the repository and install the Logfire exporter in a virtual environment:

git clone https://github.com/Kubit-AI/trace-export.git
cd trace-export
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r pydantic-logfire/requirements.txt

Then create the token file, readable only by you:

cp pydantic-logfire/.env.example pydantic-logfire/.env
chmod 600 pydantic-logfire/.env

Paste the token after LOGFIRE_READ_TOKEN= and save the file. Keep the token out of the command line and out of source files.

Check the project with a sample

Export a 10-span sample to check that the token reads the right project:

python pydantic-logfire/export_traces.py --all-time --limit 10 --output data/sample.otel.jsonl

The output names the Logfire organization and project, then the number of spans:

Output

Logfire project: <organization>/<project> (https://logfire-eu.pydantic.dev)
Downloaded 10 span rows in total
Converted 10 spans from 10 rows; skipped 0 non-span and 0 invalid rows. Output: data/sample.otel.jsonl

Without --all-time, the sample only covers the last 24 hours. If the project had no traffic in that time, you get Downloaded 0 span rows in total.

Export the traces

Export every span in a time range to one file. Set --until a few minutes in the past, because Logfire stores a span only when it ends:

python pydantic-logfire/export_traces.py --no-sample \
  --since 2026-10-01T00:00:00Z --until 2026-11-10T12:00:00Z \
  --output data/logfire.otel.jsonl --json-summary

The last line is a JSON summary. Check that status is ok and note downloaded_rows:

Output

Converted 5334 spans from 5334 rows; skipped 0 non-span and 0 invalid rows. Output: data/logfire.otel.jsonl
{"version": "1.0.0", "source": "logfire", "status": "ok", "output": "data/logfire.otel.jsonl", "format": "otel", "output_written": true, "downloaded_rows": 5334, "sample": false, "limit": null, "since": "2026-10-01T00:00:00+00:00", "until": "2026-11-10T12:00:00+00:00", "query_timeout_seconds": 300, "warnings": [], "possibly_truncated": false}

Option	What it does
`--since`, `--until`	Span start time from (inclusive) and to (exclusive), in UTC
`--environment NAME`	Only spans from that deployment environment. Repeat it for more than one.
`--limit N`	At most N spans. A limit or a time range can cut a trace in two.
`--all-time`	Every span your Logfire plan still keeps, instead of `--since` and `--until`
`--resume`	Continue an export that stopped. It needs `--keep-raw PATH` on the first run.

The file has one span per line, as OTLP JSON. It keeps every span in the range, including HTTP, database and log spans, so Kubit can show the whole trace.Constraint: The upload takes files up to 300 MB. For a larger export, use a shorter time range or --environment.

Create a Sandbox workspace

In Kubit, open the workspace menu at the top left and click New workspace. Enter a Workspace Name, keep or change the Workspace Timezone, set Origin to Sandbox and click Continue.

Create Workspace dialog with Workspace Name Logfire POC, Workspace Timezone Europe/Paris (UTC+1*) and Origin Sandbox, with the hint that it uploads OTel trace files for quick testing with no pipeline, above the Continue button.

If Sandbox isn’t in the Origin list, your organization doesn’t have it turned on. Ask support@kubit.co to turn it on.Kubit creates the workspace and opens Modeler on Upload span data.

Upload the file

Drag data/logfire.otel.jsonl onto Drop a span file here, or click browse and pick it. Then click Upload & preview.

Upload span data panel in the Modeler. A note says the first upload sets the workspace's destination to the Kubit-managed warehouse. The drop zone reads Drop a span file here or browse, JSONL or JSON up to 300 MB, from OTLP, Langfuse, Phoenix, MLflow, LangSmith, Braintrust, Datadog, AgentCore or Cortex.

The first upload sets the workspace’s destination to the Kubit-managed warehouse. You can’t switch the workspace to another connection later, even if you discard the file.The Uploads tab shows the progress in 4 steps. Keep the browser tab open until the upload finishes.

Review and import

When the upload shows Uploaded. Ready to review., Kubit reads a sample from the start of the file and shows a preview below the list of uploads. Check that Format reads OpenTelemetry (OTLP) (detected) and that Accepted spans equals Records sampled.

Preview of logfire.otel.jsonl, 55.6 MB, marked Preview ready. The file is kept until the next day. Format is OpenTelemetry (OTLP) (detected), and the tiles read Records sampled 131, Accepted spans 131, Bytes read 1 MB and Detected formats otlp-envelope · 131 records.

Further down, the preview shows:

Section	What to check
The span table	The span names you know from Logfire, such as `invoke_agent <agent name>`, `chat <model>` and `execute_tool <tool name>`
Labels in this file	Logfire exports carry no labels, so the sample has none. Leave Use these labels and scores as it is.
Proposed initial model	The model the import creates, with Fact table mapping, Derived columns and Trace structure. You don’t need to change anything.

Under After import, keep Run the Evaluator, so Kubit labels the traces as soon as the import succeeds. Tick I reviewed the sample and model preview. and click Import spans.

After import section with Run the Evaluator selected rather than I'll upload labels, the box I reviewed the sample and model preview ticked, and the Import spans and Discard buttons.

The import runs through Checking the file, Reading spans, Saving to the warehouse and Preparing reports. You can leave the page while it runs. When it’s done, the upload shows Succeeded with the number of Spans imported, which should match downloaded_rows from the export. Below it, Enrichment shows The Evaluator is labelling traces.

The upload logfire.otel.jsonl marked Succeeded, with tiles reading Spans imported 5,334, Duplicates 0, Skipped 0 and Time range Oct 30, 2026. Under Enrichment, The Evaluator is labelling traces, with Add labels file and Run again buttons.

Constraint: Kubit keeps an uploaded file for 1 day. If you don’t import it by then, upload it again.

Explore the traces

Open Activity. Keep Subject on Trace and Time Range on Last 1 Day, and click Execute. Time Range counts back from the last day in your data, shown next to it as Reports through, so older exports still show up.

Activity traces grid with a Timeline of trace counts in 5-minute bars above the columns Timestamp, Name, Spans, Latency (Ms), Tokens, Model, Labels, Trace ID, User ID and Session ID. Each row is a trace named scenario {scenario}, with its span count, latency, tokens and labels such as turn_not_advanced, low_judge_confidence, answered and repeated_tool_call. The footer reads 423 Traces, 1-50 shown.

The footer counts the traces, such as 423 Traces. Labels reads Not analyzed yet until the Evaluator has labelled that trace.Click a trace to open it. The header shows its Latency, Total cost, Tokens, User and Session. The Timeline shows the agent runs, model calls and tool calls as Pydantic AI recorded them, along with the HTTP and database spans Logfire captured. Click a chat span to see its Model, token counts and Input and Output messages.

An open trace named scenario {scenario} with an OK status, Latency 10.48s, Tokens 14,168 to 569, no User and a Session. The Timeline lists invoke_agent spans for two agents, chat gemini-3.5-flash-lite calls, an execute_tool span and a SELECT database span. The Preview of the selected chat span shows Type chat, Model gemini-3.5-flash-lite, Prompt tokens 480, Completion tokens 12 and the start of the Input messages.

An empty User means the spans carry no user ID. When the spans carry no session ID, Session shows the trace ID. To fill both for live traces, set user.id and session.id in your app. See Add user and session IDs.

Review the labels

Open Evaluator. Under Uploads, your file shows Running while the Evaluator labels its conversations, then the time of the last run.

Evaluator page titled Label your conversations. Under Uploads, logfire.otel.jsonl was imported at 6:50 PM and its last run reads Running, Labelling since 6:50 PM, next to a disabled Relabel button. Below, Questions lists the system questions Answer relevant, Task resolved and Right tools.

The Evaluator answers these system questions for each conversation:

Question	What it asks
Answer relevant	Does the assistant’s reply directly address what the user asked?
Task resolved	By the end of this conversation, had the assistant completed the task the user asked for?
Right tools	Were the tools the agent invoked the right ones for the user’s request?
Grounded in tool results	Is every identifier, amount, status and completed action in the assistant’s final reply supported by the tool results shown?

To ask your own yes or no, choice or score questions, click Add custom question, then Relabel the upload. Relabelling asks every question again and replaces the labels. The Evaluator labels one upload at a time.When the run ends, go back to Activity and click Refresh. The Labels column shows each trace’s labels, such as answered, turn_not_advanced or repeated_tool_call, as in the screenshot in Explore the traces.

Verify

The proof of concept is ready when all three are true:

The upload in Modeler > Uploads shows Succeeded, and Spans imported equals downloaded_rows in the export summary.
Activity lists as many traces as your Logfire project had in the exported range.
A trace you open shows its agent runs, chat spans with a Model and token counts, and execute_tool spans.

If it fails

Symptom	Cause	Fix
`Downloaded 0 span rows in total`	No spans in the time range. A sample without bounds covers the last 24 hours only.	Add `--all-time`, or set `--since` and `--until` to a range with traffic
The `Logfire project:` line names the wrong project	The read token belongs to another project	Create a read token in the right project and replace it in `pydantic-logfire/.env`
Warnings that end in `gen_ai_messages_not_schema_compatible`	Some messages don’t match the OpenTelemetry GenAI message format, or the app doesn’t capture message content	Nothing to fix for the upload. To capture content, don’t set `include_content=False` in `logfire.instrument_pydantic_ai()`.
Values such as session IDs read `[Scrubbed due to ...]`	Logfire scrubs values that look sensitive when it stores them	Change Logfire’s scrubbing settings for new spans. Exported spans keep the scrubbed value.
The upload is over 300 MB	The time range is too long	Export a shorter range, or one environment with `--environment`
Some traces have missing parent spans	The time range or `--limit` cut a trace	Widen the time range, and drop `--limit`

Next steps

Quickstart: send live traces to an Agent workspace.
Concepts: users, sessions, traces, spans and Polaris annotations.

Start here

Agent Observability

Product Analytics

Bring your own warehouse

Recipe: try Kubit on your Pydantic Logfire traces

Before you start

Prompt for your coding agent

Steps

Create a Logfire read token

Install trace-export

Check the project with a sample

Export the traces

Create a Sandbox workspace

Upload the file

Review and import

Explore the traces

Review the labels

Verify

If it fails

Next steps

​Before you start

​Prompt for your coding agent

​Steps

Create a Logfire read token

Install trace-export

Check the project with a sample

Export the traces

Create a Sandbox workspace

Upload the file

Review and import

Explore the traces

Review the labels

​Verify

​If it fails

​Next steps

Before you start

Prompt for your coding agent

Steps

Verify

If it fails

Next steps