OpenClaw Finally Understands OpenTelemetry APM

1Meet DataBuff: AI-native OpenTelemetry APM

DataBuff (github.com/databufflabs/databuff) is an open-source AI-native OpenTelemetry APM — deploy with one Docker command.

In one line: ingest telemetry via OTel standards first, then let AI read traces, metrics, topology, and alerts — not a chat box bolted onto dashboards.

📊 OpenTelemetry APM foundation

Native OTLP — traces, service metrics, topology, alerting
Rich data model derived from traces (services, APIs, flows, component metrics)
Three components: Ingest → Doris → Web, low ops overhead

🤖 AI-native capabilities

Multi-agent collaboration: AI brain + metrics / inspection experts
Smart query: natural language over metrics, traces, topology, alerts
Health inspection: threshold-free anomaly screening with evidence chains

Capability	What it does
In-platform AI chat	Ask in the Web UI: “Why is this trace slow?” or “Inspect order-service”
MCP for external agents (this article)	Cursor, Claude Code, OpenClaw call the same APM tools
Official Skills	Constrain query semantics and inspection flows — agents don't invent metrics

DataBuff three-component architecture — *Screenshot: Ingest → Doris → Platform — one command to run*

Global service topology — *Screenshot: global topology — AI and MCP tools read the same live data*

Multi-agent troubleshooting — *Screenshot: in-platform AI brain dispatches experts — external agents use the same backend via MCP*

This article: DataBuff exposes platform capabilities via MCP + Skills to OpenClaw (and Feishu / DingTalk channels). Below: 5-minute setup for one-sentence whole-cluster inspection.

22am alert: five tabs, or one sentence?

Alert fires: “Checkout is slow — anything wrong cluster-wide?”

Old way: APM dashboard → traces → topology → alert list → hand-written summary. 20–30 minutes per round.

Searching traces across systems — *Screenshot: multi-tab evidence stitching — agents can do it in one prompt*

With OpenClaw: send one message in CLI, WebChat, or Feishu / DingTalk. The agent calls DataBuff MCP against real OTel data in Doris, follows official Skills, and returns an evidence-backed report — not hallucinated numbers.

❌ Without MCP

The model invents “error rate 2.3%” — pure hallucination, not safe for incident channels.

✅ MCP + Skills

inspectService first → traces / topology for anomalies → Chinese or English report with evidence.

OpenClaw listing services via MCP — *Screenshot: OpenClaw activity — Databuff-apm GetCurrentTimeRange / QueryServicesAll*

3MCP is the hands; Skills are the brain

DataBuff ships an agent integration pack (integrations/agent/). You need both layers for a real ops agent:

MCP = 11 APM tools (services, traces, metrics, alerts, inspection, charts)

Skills = query semantics + inspection workflow (when to inspect, time ranges, evidence chaining)

Layer	Contents	Examples
MCP tools	Live data access	`queryServicesAll`, `inspectService`, `queryTraceDetail`
Skills	Behavior rules	`skill.data.metrics`, `skill.inspection.health`

External agents hit the same Spring beans as in-platform AI experts — not pasted APIs, real OTel data in Doris.

*Screenshot: skill.data.metrics enabled on OpenClaw Skills page*

45-minute setup: MCP + Skills end-to-end

Prerequisite: DataBuff deployed; OpenClaw can reach the Web port.

Step 0 · Get the official integration pack — do not download Skills/MCP from third-party marketplaces:

git clone https://github.com/databufflabs/databuff.git
cd databuff/integrations/agent/

Or browse integrations/agent/ on GitHub. Contains skills/ (2 official Skills) and mcp/ (client config examples).

Step 1 · Deploy DataBuff (if not installed):

curl -fsSL https://databuff.ai/databuff/ai-apm-install.sh | bash

DataBuff install success — *Screenshot: one-command install, note Web URL and port*

Step 2 · Configure MCP — copy mcp/openclaw-amc-config.example.json, set mcp.url to your DataBuff host:

{
  "name": "databuff-apm",
  "description": "DataBuff APM MCP — metrics, traces, topology, alerts, inspection",
  "mcp": {
    "transport": "streamable-http",
    "url": "http://YOUR_DATABUFF_HOST:PORT/mcp"
  },
  "skills": [
    "skill.data.metrics",
    "skill.inspection.health"
  ]
}

Replace YOUR_DATABUFF_HOST:PORT. MVP needs no Authorization header. MCP endpoint: /mcp (Streamable HTTP).

Step 3 · Install Skills — copy integrations/agent/skills/ to your client Skills path:

Client	Skills path
OpenClaw	OpenClaw workspace `skills/` (Control UI → Skills should list both)
Cursor	`~/.cursor/skills/` or project `.cursor/skills/`
Claude Code	`~/.claude/skills/`

# From databuff repo root (OpenClaw example — adjust target path)
cp -r integrations/agent/skills/skill.data.metrics   <openclaw-workspace>/skills/
cp -r integrations/agent/skills/skill.inspection.health <openclaw-workspace>/skills/

Each Skill is a directory with SKILL.md. Synced with deploy/common/skills/ — re-copy after updates.

skillId	Purpose
`skill.data.metrics`	APM metrics, trace, alert query semantics (time ranges, which tool)
`skill.inspection.health`	Health inspection flow (inspect first, gather evidence, conclude)

Step 4 · Verify — after restarting OpenClaw Gateway:

Control UI → Skills: skill.data.metrics, skill.inspection.health
Control UI → Instances: Gateway online
New chat → Activity: Databuff-apm tools (e.g. QueryServicesAll)

OpenClaw connected instances — *Screenshot: Gateway and Control UI connected (127.0.0.1:18789)*

Security: MVP has no separate MCP token — use on private network / VPN; add a gateway before public exposure.

5Three prompts to validate the full path

①

“List services from the last hour” — Skill parses time → getCurrentTimeRange → queryServicesAll

List services MCP calls — *Screenshot: 7 services in the last hour (service-a / service-b / MySQL, etc.)*

②

“Inspect service-a health” — inspectService → queryMetricData / queryServiceAlarms

Inspect service-a tool chain — *Screenshot: Activity 3 tools — InspectService / QueryMetricData / QueryServiceAlarms*

③

“Chart service-a error rate for the past hour” — queryMetricData + drawTrendCharts

Error rate trend chart — *Screenshot: 0% error rate trend — real data, not invented*

Pass criteria: tool panel shows queryServicesAll / inspectService with JSON responses — not plain-text guesses.

6vs in-platform AI and vs Cursor?

Scenario	Path	Best for
DataBuff Web UI	AgentScope JAVA_BEAN + multi-expert	Day-to-day console troubleshooting
OpenClaw / Feishu / DingTalk	MCP + Skills	One-liner queries in alert channels
Cursor / Claude Code	Same MCP + Skills	Developers querying APM from the IDE

All three paths are independent. OpenClaw brings APM into your existing agent runtime — no extra browser tab for AI incidents.

Full config and FAQ: docs/manual/agent-integration.md and integrations/agent/README.md in the repo.

7From guessing metrics to calling real APM

DataBuff opens APM to any agent runtime; OpenClaw / Feishu / DingTalk are just entry points — MCP for real data, Skills for semantics is the core.

⭐ Wire up DataBuff agent integration in 5 minutes

github.com/databufflabs/databuff/integrations/agent

Try: “Inspect service-a, then check if anything is wrong cluster-wide.”

View on GitHub →