Cloudflare AI Gateway

Cloudflare AI Gateway routes requests to model providers and provides caching, rate limiting, retries, and analytics. Its native OpenTelemetry exporter can send gateway traces directly to Parseable, where you can analyze requests, errors, latency, token usage, cost, providers, and models.

How it works

Application
  |
  v
Cloudflare AI Gateway ----> Model provider
  |
  | OTLP/HTTP traces over HTTPS
  v
Parseable /v1/traces
  |
  v
cf-ai-gateway-traces

Prerequisites

• A Cloudflare account with an AI Gateway
• A Cloudflare API token with AI Gateway Read and Edit permissions
• A model-provider key, Cloudflare unified billing, or Workers AI access
• A running Parseable instance with a publicly reachable HTTPS endpoint
• A Parseable API key with ingest access

Set up Cloudflare AI Gateway with Parseable

export PARSEABLE_URL="https://parseable.example.com"
export PARSEABLE_API_KEY="<parseable-api-key>"
export PARSEABLE_STREAM="cf-ai-gateway-traces"

curl -X PUT "$PARSEABLE_URL/api/v1/logstream/$PARSEABLE_STREAM" \
  -H "X-API-Key: ${PARSEABLE_API_KEY}" \
  -H "X-P-Log-Source: otel-traces" \
  -H "X-P-Telemetry-Type: traces"

cloudflared tunnel --url http://localhost:8010

export CLOUDFLARE_ACCOUNT_ID="<account-id>"
export CLOUDFLARE_GATEWAY_ID="<gateway-id>"
export CLOUDFLARE_AIG_TOKEN="<authenticated-gateway-token>"

curl -X POST \
  "https://gateway.ai.cloudflare.com/v1/$CLOUDFLARE_ACCOUNT_ID/$CLOUDFLARE_GATEWAY_ID/openai/chat/completions" \
  -H "cf-aig-authorization: Bearer ${CLOUDFLARE_AIG_TOKEN}" \
  -H "Content-Type: application/json" \
  -H 'cf-aig-metadata: {"workload":"validation","environment":"development"}' \
  -d '{
    "model": "gpt-4.1-mini",
    "messages": [
      {
        "role": "user",
        "content": "Explain observability in one sentence."
      }
    ],
    "max_tokens": 100
  }'

What you get in Parseable

Open cf-ai-gateway-traces from the Traces page. Cloudflare emits standard GenAI span attributes and converts values from cf-aig-metadata into searchable span attributes.

Common fields include:

Prompt and completion fields can contain sensitive data. Review Cloudflare logging controls and your retention policy before enabling this integration in production.

Verify ingestion

In Parseable's SQL editor, select cf-ai-gateway-traces and run:

SELECT
  p_timestamp,
  span_trace_id,
  "gen_ai.model.provider",
  "gen_ai.request.model",
  "gen_ai.usage.input_tokens",
  "gen_ai.usage.output_tokens",
  "gen_ai.usage.cost"
FROM "cf-ai-gateway-traces"
ORDER BY p_timestamp DESC
LIMIT 20;

To summarize traffic by provider and model:

SELECT
  "gen_ai.model.provider" AS provider,
  "gen_ai.request.model" AS model,
  COUNT(*) AS requests,
  SUM(COALESCE("gen_ai.usage.input_tokens", 0)) AS input_tokens,
  SUM(COALESCE("gen_ai.usage.output_tokens", 0)) AS output_tokens,
  SUM(COALESCE("gen_ai.usage.cost", 0)) AS cost
FROM "cf-ai-gateway-traces"
GROUP BY provider, model
ORDER BY requests DESC;

Dashboard template

Parseable provides a ready-to-import Cloudflare AI Gateway Observability dashboard. It contains 32 SQL-backed tiles across six sections:

• Overview
• Traffic and reliability
• Performance and latency
• Tokens and cost
• Models and workloads
• Trace explorer

Download the dashboard JSON template, then open Dashboards in Parseable and use the dashboard import flow. Map Trace Dataset to cf-ai-gateway-traces during import or after the dashboard is created.

The template uses native Cloudflare OTLP spans and requires no separate metrics dataset. Custom metadata fields such as workload, complexity, and source improve its breakdowns but are optional. After importing, select a time range that contains your gateway traffic.

Add custom metadata and trace context

Attach JSON metadata to a request with cf-aig-metadata. Use stable dimensions such as workload, environment, tenant, or team so dashboards can group traffic without creating excessive cardinality.

To connect gateway calls to an existing distributed trace, Cloudflare accepts these headers:

• cf-aig-otel-trace-id: a 32-character hexadecimal trace ID
• cf-aig-otel-parent-span-id: a 16-character hexadecimal parent span ID

Troubleshooting

• No traces arrive: Confirm that the exporter URL is public HTTPS and ends in /v1/traces. If you are using a Quick Tunnel, make sure it is still running and that the current URL matches the exporter configuration. Check that the request passed through the same Cloudflare gateway whose exporter you configured, then generate a new request and expand the Parseable time range.
• Parseable returns an authentication error: Verify X-API-Key. When using that header, leave Cloudflare's exporter Authorization field empty.
• Parseable reports a missing stream or log source: Create cf-ai-gateway-traces, set X-P-Stream to that exact value, and set X-P-Log-Source to otel-traces.
• The model request is rejected: Verify the account ID, gateway ID, cf-aig-authorization token, stored provider key, model name, and provider plan. Some models are unavailable on free plans.
• Dashboard charts show no data: Choose a time range containing ingested traces and confirm that the dashboard dataset variable is cf-ai-gateway-traces. Cost, token, or latency charts remain empty when the selected traces do not contain those attributes.

References

• Cloudflare OpenTelemetry integration
• Cloudflare AI Gateway setup
• Cloudflare bring-your-own-key setup