Git is the source of truth
The catalogue is bundled at build time, so every merge deploys and a broken skill breaks the build instead of reaching an agent.
SRE tooling · 2026 · Agent Skills and MCP
Runbooks an agent can follow without improvising
A versioned catalogue of SRE and Google Cloud skills, with the investigation patterns and scripts they call, served to Claude Code, claude.ai, a LangGraph incident agent and local models through one private MCP server.
- 30
- skills: 25 general, 5 for incident investigation
- 77
- reference runbooks, loaded only when needed
- 28
- investigation patterns that classify signal from noise
- 34
- scripts the skills call instead of guessing
The problem
An LLM asked to investigate an incident will happily improvise: a plausible query, a confident root cause, a symptom mistaken for the cause. In operations that is worse than no answer.
The fix is to move the know-how out of the prompt and into versioned files. The model chooses which skill applies; the skill decides the steps, the queries and what counts as evidence. Same alert, same procedure, every time.
How it works
From commit to every agent
- WriteSKILL.md with a trigger description, references and scripts, in git
- ValidateBuild fails if a name differs from its folder, a description is empty or a name repeats
- BundleCatalogue compiled into the Worker on every deploy
- ServeFour MCP tools: list_skills, get_skill, get_reference, search_skills
- ProtectOAuth with owner-only consent, or a bearer token; nothing public
- ConsumeClaude Code, claude.ai, the incident agent and Ollama models
What a pattern encodes
Investigation skills are rules, not prompts. A few of the principles every incident pattern follows:
investigation-skills/incident-investigation/SKILL.md
## Principles
1. Filter noise before root cause analysis.
Health probes and expected payment
declines are not incidents.
2. Root cause precedes symptom in time.
OOM precedes liveness failure.
3. Alert direction encodes the problem.
GT = consumer failing,
LT = publisher stopped.
4. Return structured JSON only.Decisions that mattered
Generic skills, tenant values apart
No company names, project IDs or hostnames in skills. Employer-specific values live under tenants/ and are never served.
Small context first
References are opt-in, so the same skills work in a 32k context on a local model and in a frontier model.
Private by default
Read-only server, owner-only OAuth consent, constant-time token checks, and a 503 if the secret is missing.
Scripts over prose
Fetching logs, time series, revisions or feature-flag changes is a function with a typed return, not an instruction to the model.
Validate where the rules live
Drafting happens here; a handoff skill produces the prompt that validates the result inside each company's own assistant.
Stack
| Layer | Choice | Why |
|---|---|---|
| Content | Agent Skills (Markdown) | Readable by people and by any compatible client |
| Scripts | Python, Google Cloud APIs | Logs without a severity filter, metrics, revisions, SLOs |
| Server | TypeScript, MCP SDK, Cloudflare Workers | Stateless Streamable HTTP, one server per request |
| Auth | workers-oauth-provider + KV | Dynamic client registration for claude.ai |
| Tests | node --test | End to end against the production origin |
Where it stands
In daily use by the incident agent and in day-to-day SRE work. Started as a fork of Jeffallan/claude-skills (MIT); the upstream skills kept carry their original attribution, and the investigation skills, patterns, scripts and MCP server are new.
More work
Production RAG · WhatsAppGenerative AI in production, on your own infrastructureClient work · ML + fraudWho buys parts next month, and which invoice is a fraudAgentic SRE · LangGraphFrom a vague alert to an evidence-backed hypothesisSide project · Sports analyticsTraining analysis a coach can defend
Have an AI project stuck at the prototype stage?
Tell us the problem, the data and where it needs to run. We'll come back with an architecture and scope proposal.
Email contato@laloop.ai