Logfire plugin for Claude Code
A Claude Code plugin that sends OpenTelemetry traces to Pydantic Logfire, giving you full observability into your Claude Code sessions.
Each session becomes a trace with child spans per LLM API call, with full token usage, cost tracking, and conversation history visible in Logfire.
Installation
System requirements
- Claude Code installed
- A Logfire project with a write token
python3 (3.7+) — pre-installed on macOS and most Linux distributions; uses stdlib only (no pip dependencies)
Install the plugin
From within Claude Code, run:
/plugin marketplace add pydantic/claude-code-logfire-plugin
/plugin install logfire-session-capture@pydantic-claude-code-logfire-plugin
Update the plugin
/plugin update logfire-session-capture@pydantic-claude-code-logfire-plugin
Set your Logfire token
export LOGFIRE_TOKEN="your-logfire-write-token"
Add this to your shell profile (~/.zshrc, ~/.bashrc, etc.) so it persists across sessions.
For the EU region:
export LOGFIRE_BASE_URL="https://logfire-eu.pydantic.dev"
| Variable | Required | Default | Description |
|---|
LOGFIRE_TOKEN | Yes | (none) | Logfire write token |
LOGFIRE_BASE_URL | No | https://logfire-us.pydantic.dev | Logfire ingest endpoint |
LOGFIRE_LOCAL_LOG | No | false | Set to true to write JSONL event logs locally |
LOGFIRE_DIAGNOSTICS | No | false | Set to true to write diagnostic logs (enabled automatically when LOGFIRE_LOCAL_LOG is set) |
LOGFIRE_ENVIRONMENT | No | (none) | Sets deployment.environment.name on every trace (e.g. production, dev) |
LOGFIRE_SESSION_LABEL | No | Claude Code session | Label for the root session span — useful when running multiple CC sessions in one trace |
OTEL_SERVICE_NAME | No | claude-code-plugin | Overrides the service.name resource attribute |
OTEL_RESOURCE_ATTRIBUTES | No | (none) | Standard OTel env var for additional resource attributes, e.g. deployment.environment.name=prod,service.instance.id=worker-1 |
Without LOGFIRE_TOKEN, no traces are sent. The plugin does nothing unless at least one of LOGFIRE_TOKEN or LOGFIRE_LOCAL_LOG is set.
What you get
Every Claude Code session produces a trace in Logfire:
Claude Code session <- root span (the full session)
├── chat claude-opus-4-6 <- LLM API call 1
├── chat claude-opus-4-6 <- LLM API call 2
└── chat claude-opus-4-6 <- LLM API call 3
Each chat child span includes:
- Token usage (
gen_ai.usage.input_tokens, gen_ai.usage.output_tokens)
- Cost (
operation.cost in USD)
- Messages (
gen_ai.input.messages, gen_ai.output.messages)
- Finish reason (
gen_ai.response.finish_reasons)
The root span carries the full conversation, so you can inspect the entire session in Logfire's trace view.
Distributed tracing
If you call Claude Code from a Python application that already uses Logfire or OpenTelemetry, you can link the Claude Code session into your existing trace by passing a TRACEPARENT environment variable:
TRACEPARENT="00-<trace_id>-<parent_span_id>-01" claude --print "your prompt"
See examples/distributed-tracing.py for a complete example using logfire and subprocess.
Local JSONL log
Set LOGFIRE_LOCAL_LOG=true to write all hook events as JSON Lines to .claude/logs/session-events.jsonl in the project directory. This is off by default.
Data collected
When LOGFIRE_TOKEN is set, the plugin sends the following data to Logfire as OpenTelemetry span attributes:
| Data | Span | Attribute |
|---|
| Full conversation (user prompts, assistant responses, tool calls and results) | Root span | pydantic_ai.all_messages |
| Per-call input/output messages | Child spans | gen_ai.input.messages, gen_ai.output.messages |
| Token counts | Child spans | gen_ai.usage.input_tokens, gen_ai.usage.output_tokens |
| Cost in USD | Child spans | operation.cost |
| Model name | Both | gen_ai.request.model |
| Working directory | Root span | session.cwd |
| Assistant thinking blocks | Child spans | Included in gen_ai.output.messages |
Privacy note: Conversation data sent to Logfire may contain sensitive information including file contents read by Claude, tool outputs, environment details, and any text in the conversation. Logfire data is stored according to Pydantic's privacy policy. If this is a concern, use LOGFIRE_LOCAL_LOG=true without LOGFIRE_TOKEN to keep all data local.
Troubleshooting
Enable diagnostics to see what the plugin is doing:
export LOGFIRE_DIAGNOSTICS=true
