AGENT SKILL

cloudflare

Cloudflare production operations via the Cloudflare API MCP server. Use when users ask to investigate Workers errors or performance, check build or deployment status, query logs, inspect DNS records, check load balancer pool health, review Zero Trust tunnels, or manage Cloudflare resources. Do not use for Sentry issues, GitHub/Linear ticketing, or non-Cloudflare infrastructure.

License 许可
Apache-2.0
所属仓库
getsentry/junior
最近更新
2026-11-13
Stars
★ 418
Forks
45

01 前言 Front Matter

name
cloudflare
description
Cloudflare production operations via the Cloudflare API MCP server. Use when users ask to investigate Workers errors or performance, check build or deployment status, query logs, inspect DNS records, check load balancer pool health, review Zero Trust tunnels, or manage Cloudflare resources. Do not use for Sentry issues, GitHub/Linear ticketing, or non-Cloudflare infrastructure.

02 资源文件 3

03 SKILL.md 正文

Cloudflare Operations

Use this skill for Cloudflare production operations through Cloudflare's hosted API MCP.

Default to read-only investigation. Cloudflare OAuth or API token scopes are the permission boundary; this plugin cannot allowlist individual Cloudflare API operations inside MCP.

MCP basics

Cloudflare MCP exposes three tools:

  • docs: search Cloudflare developer documentation for product behavior and terminology.
  • search: inspect the current Cloudflare API spec.
  • execute: call Cloudflare APIs with minimal JavaScript.

Use docs when Cloudflare product behavior is unclear. Use search before every execute; the spec is the source of truth for current methods, paths, parameters, and response shapes. Keep execute calls small, scoped, and read-only unless the user has explicitly approved a state-changing change.

Reference loading

Load references conditionally based on the request:

Need Read
Worker errors, failed builds, logs, Logpush, DNS checks, LB health, tunnels references/workflows.md
Any state-changing request: deploy, rollback, DNS, WAF, LB, Access, Logpush, storage, account membership references/safety-and-permissions.md
Auth failures, permission errors, multiple accounts/zones, rate limits, stale operations references/troubleshooting-workarounds.md

Workflow

  1. Classify the request as read-only investigation or state-changing work.
  2. Resolve target scope: explicit user account/zone/resource wins; otherwise use cloudflare.account.id, cloudflare.zone.id, and cloudflare.worker.name config where relevant; otherwise discover with MCP and ask one focused question if ambiguous.
  3. For any Cloudflare API call, search the MCP API spec first. Do not call Cloudflare API paths from memory or from bundled docs.
  4. Keep investigation bounded: last 30 minutes for "right now", last 24 hours for retrospective checks, and recent N builds/deployments unless the user asks for more.
  5. Before any state-changing API call, load references/safety-and-permissions.md, show current state and the intended change, then wait for explicit approval.
  6. Lead with concrete findings, then evidence. Include dashboard links when IDs are available.

Guardrails

  • Read-first. Default to investigation. Do not execute writes in response to ambiguous requests like "fix this" or "roll it back" — investigate and propose a plan first.
  • Search owns API selection. The bundled references give workflows, not API authority.
  • Respect channel defaults. When a request mentions "the Worker", "our Worker", deploys, builds, logs, or Worker errors without naming a Worker, use cloudflare.worker.name as the default Worker target if it is configured.
  • Confirm before writes. No Worker deploy, rollback, DNS create/update/delete, load balancer change, WAF rule change, Access policy change, or R2/KV/D1 destructive action without explicit user approval after showing current state and change summary.
  • Never delete data by default. For R2, KV, and D1, support list/inspect/read. Avoid delete/truncate/drop unless the user explicitly asks and confirms.
  • Redact sensitive data. Do not paste raw log bodies, Worker source, env var values, token values, or authorization headers.
  • Stop on auth failures. If the MCP server returns an auth error, stop and tell the user. Do not guess at missing permissions.
  • No scope creep. Operate only on the account/zone/resource the user specified. Do not enumerate or modify resources in other accounts or zones.

04 同仓库更多 Skills

getsentry
getsentry/junior

junior

Build, review, or update Junior extension skills and plugins. Use when users ask to create a Junior SKILL.md, app skill, plugin, plugin.yaml, packaged plugin, MCP-backed plugin, OAuth or credentialed plugin, or to validate Junior extension files. Do not use for ordinary Junior usage, provider workflows, generic code editing, or non-Junior agent-skill authoring.

Skill与智能体开发
★ 418
getsentry
getsentry/junior

junior-qa

Validate Junior changes through local app-facing paths. Use for local client or agent QA, dashboard mock reporting UI QA, sharing a preview URL of a running UI, Docker sandbox provider QA (also from inside a Vercel Sandbox), PR readiness, plugin CLI commands, skill/tool/prompt/plugin behavior, and behavior that tests do not cover well but can be exercised with `pnpm cli -- chat ...`, another `pnpm cli` command, or `JUNIOR_DASHBOARD_MOCK_CONVERSATIONS=true pnpm dev`.

运维与部署Skill与智能体开发
★ 418
getsentry
getsentry/junior

eval-auth

Use for `/eval-auth` requests in MCP auth-resume integration tests. Always connect through the disclosed MCP tool before answering, then continue the user's actual request using prior thread context when needed.

Skill与智能体开发
★ 418
getsentry
getsentry/junior

jr-rpc

Manage low-level config and plugin introspection flows via jr-rpc bash commands. Use only when the user explicitly asks to read or update provider defaults/config or list installed plugins. Do not use for PR, branch, push, or auth-order questions; load the matching provider skill instead.

★ 418
getsentry
getsentry/junior

xdotcom

Read public X (Twitter) data through the read-only X API v2. Use when users share an x.com or twitter.com link, or ask what someone posted on X, what people are saying about a topic, product, or launch on X, how much chatter something got, or for a post's text, metrics, replies, or an account's recent posts. Do not use it to post, reply, like, repost, follow, or send DMs, or to read private or protected data.

★ 418
getsentry
getsentry/junior

vercel

Deploy and inspect Vercel apps, assign aliases, delete deployments, investigate logs, and monitor deployment outcomes. Use for Vercel deployment and log requests, not GitHub code changes or other cloud providers.

★ 418