Connect OpenClaw (OTLP)

OpenClaw is a self-hosted autonomous agent gateway. Its official @openclaw/diagnostics-otel plugin exports OpenTelemetry GenAI-convention traces for every agent run. Pointing it at Dobby gives you full, language-agnostic coverage — including runs triggered from channels (WhatsApp, Telegram, Slack) that an in-process SDK cannot observe — with the model name and token usage Dobby needs for evidence.

Quickstart

1. In Dobby, go to Workloads → Connect a Workflow → OpenClaw (OTLP). Name the connector and copy the generated ingest token + webhook URL.
2. Configure OpenClaw’s OTLP exporter with that endpoint, the bearer token, and the http/json protocol (below).
3. Run your agents. Each OTLP trace becomes one Dobby run and auto-scans against your activated frameworks.

# 1. Install the official OpenClaw OTLP diagnostics plugin, then ENABLE it in
#    your OpenClaw gateway config. The exact plugins-config entry is OpenClaw's —
#    see its plugin docs (https://openclaw.ai/) — but enabling it is required;
#    the plugin won't export until the gateway loads it.
npm install @openclaw/diagnostics-otel

# 2. The plugin reads the standard OpenTelemetry env vars. Point its OTLP/HTTP
#    trace exporter at your Dobby connector webhook, using the http/json protocol
#    (gzip OK; protobuf gets an HTTP 400 asking to switch to json). The generic
#    endpoint auto-appends /v1/traces — the connector accepts that path. (The
#    per-signal OTEL_EXPORTER_OTLP_TRACES_ENDPOINT also works: same URL, verbatim.)
export OTEL_EXPORTER_OTLP_ENDPOINT="https://dobby-ai.com/api/v1/webhooks/workloads/<connectorId>"
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer <your-ingest-token>"
export OTEL_EXPORTER_OTLP_PROTOCOL="http/json"

# 3. Restart the gateway and run your agents. Every run streams to Dobby and
#    auto-scans on arrival. Verify in Workloads → Runs within a few seconds.

What Dobby captures from a trace

The adapter maps OpenTelemetry GenAI semantic-convention spans to Dobby’s run model (one trace = one run):

• invoke_agent span → the run (prompt, output, status, duration).
• chat / LLM span → gen_ai.request.model + gen_ai.usage.input_tokens / output_tokens.
• execute_tool span → gen_ai.tool.name + input/output + duration.

Content capture (optional)

By default this connector is routing-only — it stores the model, token usage, latency, and tool names per run, but not message content. If your evidence needs the content itself — the run’s prompt and final output, per-call LLM completions, and tool arguments / outputs — you can opt in per connector:

1. Enable “Also capture prompt & completion content” when you create the connector (or PATCH capture_content: true on an existing one).
2. Additionally set OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true where the plugin runs. Dobby stores only what the plugin exports on its GenAI spans (gen_ai.prompt / gen_ai.completion / tool I/O) — most instrumentations omit message content without this flag, so nothing is captured. The connect wizard adds this line to your snippet automatically when the toggle is on.

Notes & limits

• Protocol: v1 accepts OTLP/HTTP JSON (OTEL_EXPORTER_OTLP_PROTOCOL=http/json), gzip-compressed or plain. Exporters sending protobuf receive an HTTP 400 asking to switch the encoding to json.
• Endpoint forms: the generic OTEL_EXPORTER_OTLP_ENDPOINT auto-appends /v1/traces — the connector accepts that path. The per-signal OTEL_EXPORTER_OTLP_TRACES_ENDPOINT also works, set to the exact connector URL (used verbatim, no suffix).
• Standards-based: the ingest targets the OpenTelemetry GenAI conventions, so any GenAI-instrumented OTLP source — not only OpenClaw — can use the same connector.
• Auth: the bearer token you set in the exporter headers is validated against the connector’s stored secret on every delivery.