beyla — Claude Code skill: what it does, how to install
Auto-instruments HTTP, gRPC, and database traffic using Grafana Beyla eBPF without application code changes. Useful when you need traces and RED metrics from prebuilt binaries.
npx skills add grafana/skills --skill beyla -a claude-codeWhat beyla does
This skill guides the configuration of Grafana Beyla to inspect application traffic across Go, Java, Python, Node, Ruby, .NET, Rust, C++, and PHP. It provides workflows for running Beyla in Docker alongside an existing process or rolling it out as a cluster-wide Kubernetes DaemonSet. Beyla captures runtime network calls, forwarding OTLP traces to Tempo and Prometheus metrics to scraping endpoints.
System prerequisites are strict. Without Linux kernel 5.8 or higher with BTF enabled, Beyla fails to load its BPF objects. Containers also require CAP_SYS_ADMIN or privileged: true paired with hostPID: true to inspect host sockets. The skill includes setup patterns for trace sampling, Grafana Cloud routing via Alloy, and heuristic route decorators to stop Prometheus cardinality explosion.
What it helps with
- collect OTLP traces and RED metrics without modifying source code
- deploy Grafana Beyla as a cluster-wide Kubernetes DaemonSet
- prevent Prometheus route cardinality explosions using heuristic decorators
- forward eBPF telemetry to Grafana Cloud through Alloy
Example prompts for the agent
- Help me run Grafana Beyla in Docker against a service listening on port 8080.
- Generate a Kubernetes DaemonSet manifest for Beyla exporting to Tempo.
- How do I fix the failed to load BPF object error when launching Beyla?
How to install beyla
1Skills CLI (Vercel)
npx skills add grafana/skills --skill beyla -a claude-code2Manually: put the files in place
~/.claude/skills/beyla/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/beyla/
1Skills CLI (Vercel)
npx skills add grafana/skills --skill beyla -a codex2Manually: put the files in place
.agents/skills/beyla/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add grafana/skills --skill beyla -a cursor2Manually: put the files in place
.cursor/skills/beyla/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install grafana/skills beyla2Manually: put the files in place
.github/skills/beyla/SKILL.md1Skills CLI (Vercel)
npx skills add grafana/skills --skill beyla -a gemini-cli2Platform CLI
gemini skills install https://github.com/grafana/skills --path skills/grafana-core/beylaManually: put the files in place
.devin/skills/beyla/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add grafana/skills --skill beyla -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about beyla
How to install the beyla skill in Claude Code?
Run npx skills add grafana/skills --skill beyla -a claude-code in your terminal. This registers the skill for Claude Code.
What does the beyla skill do?
It provides ready-to-use workflows for deploying Grafana Beyla via Docker or a Kubernetes DaemonSet. It helps configure zero-code eBPF tracing and Prometheus metrics collection for backend services.
What prerequisites are required to run Beyla?
You need a Linux system on kernel 5.8 or newer with BTF enabled (/sys/kernel/btf/vmlinux), an x86_64 or ARM64 architecture, and root permissions or CAP_SYS_ADMIN with hostPID: true.
Source text
SKILL.md · Apache-2.0
---
name: beyla
license: Apache-2.0
description: Auto-instrument an application's HTTP / gRPC / DB traffic with Grafana Beyla eBPF — no code changes, no SDK, no restart. Covers requirements (Linux 5.8+ with BTF, CAP_SYS_ADMIN, host PID), language matrix (Go / Java / Python / Ruby / Node / .NET / Rust / C++ / PHP), Docker + Helm + DaemonSet install, port- / process- / Kubernetes-metadata discovery, OTLP traces + Prometheus metrics export, routes decorator (cardinality control), trace sampling, and Grafana Cloud via Alloy. Use when adding observability to a service you can't recompile, instrumenting a closed-source binary, getting RED metrics + spans onto Tempo/Mimir without touching the app, or rolling Beyla as a cluster-wide DaemonSet — even when the user says "zero-code APM", "instrument legacy app", "trace this binary", "eBPF observability", or "no SDK" without naming Beyla.
---
# Grafana Beyla
> **Docs**: https://grafana.com/docs/beyla/latest/
Zero-code HTTP / gRPC / DB instrumentation via eBPF. Emits OTLP traces + Prometheus metrics.
## Prerequisites
- Linux kernel **5.8+** with BTF enabled (`ls /sys/kernel/btf/vmlinux` must exist)
- Root or `CAP_SYS_ADMIN` (or `privileged: true` in Kubernetes + `hostPID: true`)
- x86_64 or ARM64
- An OTLP receiver (Tempo, Alloy, OTel Collector) reachable from Beyla
## Common Workflows
### 1. Instrument a single binary with Docker
```bash
# 1. Run Beyla against the app's port (the app must already be running, listening on 8080)
docker run --privileged --pid=host \
-v /sys/kernel/debug:/sys/kernel/debug:ro \
-e BEYLA_OPEN_PORT=8080 \
-e BEYLA_PROMETHEUS_PORT=8999 \
-e OTEL_EXPORTER_OTLP_ENDPOINT=http://otel-collector:4318 \
-p 8999:8999 \
grafana/beyla
# 2. Generate some traffic
curl http://localhost:8080/ ; curl http://localhost:8080/api/users/42
# 3. Verify Beyla emitted metrics — should list http_server_request_duration_seconds + counters
curl -s http://localhost:8999/metrics | grep -E '^http_(server|client)_request_duration'
# 4. Verify traces — in Grafana Explore on Tempo, search by service.name (default = process name)
# Or: query Tempo's search API for spans with service.name="<app>"
```
### 2. Deploy as a cluster-wide DaemonSet
Full DaemonSet + RBAC YAML lives in [`references/kubernetes.md`](references/kubernetes.md). After applying:
```bash
# 1. Verify DaemonSet rollout
kubectl -n monitoring rollout status ds/beyla
# 2. Verify pods are Running, one per node
kubectl -n monitoring get pods -l app=beyla -o wide
# 3. Verify eBPF probes attached (no errors mentioning BTF or "permission denied")
kubectl -n monitoring logs ds/beyla --tail=50 | grep -Ei 'error|fail|btf' || echo "clean"
# 4. Verify telemetry is flowing — check the Tempo/Alloy receiver for spans from the cluster
# Or scrape one pod directly:
kubectl -n monitoring port-forward ds/beyla 8999:8999 &
curl -s localhost:8999/metrics | head
```
### 3. Send to Grafana Cloud via Alloy
```yaml
# beyla-config.yml
otel_traces_export: { endpoint: http://alloy:4318 }
otel_metrics_export: { endpoint: http://alloy:4318 }
```
```bash
# Verify Alloy is forwarding — check Alloy UI (localhost:12345) for the
# otelcol.receiver.otlp.beyla component showing received spans/metrics > 0.
```
Full Alloy + Beyla YAML: [`references/config.md`](references/config.md).
## Troubleshooting
- `failed to load BPF object` → kernel < 5.8 or BTF missing; check `/sys/kernel/btf/vmlinux`
- No spans in Tempo, but Prometheus metrics show → check `OTEL_EXPORTER_OTLP_ENDPOINT`, protocol (http vs grpc), and ports (4318 http / 4317 grpc)
- HTTP route cardinality explosion → set `routes.unmatched: heuristic` and add patterns (see [`references/config.md`](references/config.md))
- Pod restarts with `CrashLoopBackOff` → likely missing `hostPID: true` or `privileged: true` / required capabilities
## Resources
- [Beyla docs](https://grafana.com/docs/beyla/latest/)
- [Beyla GitHub](https://github.com/grafana/beyla)
- [`references/config.md`](references/config.md) — full config, env vars, samplers, routes decorator, generated metrics table, runtime matrix
- [`references/kubernetes.md`](references/kubernetes.md) — DaemonSet + RBAC + discovery filters + Helm
© grafana · Apache-2.0 · grafana/skills