logfire-observability

Full OpenClaw observability in Pydantic Logfire. Get agent traces, tool calls, metrics, and logs — all in one dashboard.

This setup combines two plugins:

Plugin	What it sends to Logfire	Source
logfire-observability (this plugin)	Agent→tool trace hierarchy with params, results, and parent-child nesting	Custom, ships here
diagnostics-otel (built-in)	Metrics (tokens, cost, duration), logs, webhook/queue/session telemetry	Ships with OpenClaw

Both are configured to export to Logfire via OTLP. Together they give you full coverage.

Quick start

1. Get a Logfire token

Go to logfire.pydantic.dev
Create a project (or use an existing one)
Go to Settings > Write Tokens > Create Token
Copy the pylf_v1_us_... token

2. Install this plugin

openclaw plugins install openclaw-logfire-observability

That's it. OpenClaw downloads the plugin from npm and wires it into your extensions.

Alternative: install from source

git clone https://github.com/rita-aga/openclaw-logfire-observability.git ~/.openclaw/extensions/logfire-observability
cd ~/.openclaw/extensions/logfire-observability && npm install

Or add it as a git submodule in your project and link it:

openclaw plugins install -l /path/to/openclaw-plugins/logfire-observability

3. Configure both plugins

Add the following to your openclaw.json (or ~/.clawdbot/openclaw.json).

Replace YOUR_TOKEN_HERE with your Logfire write token in both places:

{
  "diagnostics": {
    "enabled": true,
    "otel": {
      "enabled": true,
      "endpoint": "https://logfire-us.pydantic.dev",
      "headers": {
        "Authorization": "Bearer pylf_v1_us_YOUR_TOKEN_HERE"
      },
      "serviceName": "openclaw",
      "traces": true,
      "metrics": true,
      "logs": true
    }
  },
  "plugins": {
    "entries": {
      "openclaw-logfire-observability": {
        "enabled": true,
        "config": {
          "logfireToken": "pylf_v1_us_YOUR_TOKEN_HERE"
        }
      },
      "diagnostics-otel": {
        "enabled": true
      }
    }
  }
}

Then restart OpenClaw (sudo systemctl restart clawdbot or openclaw restart).

EU region? Change the endpoint to https://logfire-eu.pydantic.dev and set logfireEndpoint to https://logfire-eu.pydantic.dev/v1/traces.

What you get in Logfire

From logfire-observability (this plugin)

Detailed agent execution traces with parent-child nesting:

User message
  └─ message.received span
  └─ agent.run span (parent)
       ├─ tool.web_search span
       ├─ tool.read_file span
       └─ tool.send_message span

Span	Fires when	Key attributes
`message.received`	Inbound user message	channel, from, content
`agent.run`	LLM call start → end	agent, provider, prompt preview, response, duration, message count, token usage, cost, model
`tool.<name>`	Each tool execution	tool name, params, result, call ID

All spans include openclaw.sessionKey and openclaw.agent for filtering.

agent.run spans also include OpenTelemetry GenAI semantic convention attributes:

Attribute	Description
`gen_ai.usage.input_tokens`	Total input tokens
`gen_ai.usage.output_tokens`	Total output tokens
`gen_ai.usage.total_tokens`	Combined total
`gen_ai.usage.cache_read_tokens`	Tokens served from cache
`gen_ai.usage.cache_write_tokens`	Tokens written to cache
`gen_ai.response.model`	Model used for the response
`openclaw.llm.cost_usd`	Estimated cost in USD

From diagnostics-otel (built-in)

Operational metrics, logs, and diagnostic traces:

Metrics

Metric	Type	What it tracks
`openclaw.tokens`	counter	Token usage by type (input, output, cache, prompt, total)
`openclaw.cost.usd`	counter	Estimated cost per run
`openclaw.run.duration_ms`	histogram	Agent run duration
`openclaw.context.tokens`	histogram	Context window limit vs used
`openclaw.webhook.received`	counter	Inbound webhooks
`openclaw.webhook.duration_ms`	histogram	Webhook processing time
`openclaw.message.queued` / `.processed`	counters	Message throughput
`openclaw.queue.depth` / `.wait_ms`	histograms	Queue health
`openclaw.session.state` / `.stuck`	counters	Session lifecycle
`openclaw.run.attempt`	counter	Run retry tracking

Logs — All OpenClaw logs forwarded to Logfire via OTLP (when logs: true).

Traces — model.usage, webhook.processed, webhook.error, message.processed, session.stuck spans.

Config reference