Cloudflare MCP — Сценарии использования, Установка & Живое демо

Name: Cloudflare MCP Server
Author: cloudflare

Зачем использовать

Ключевые функции

Remote MCP — hosted by Cloudflare, OAuth sign-in, no local server process
Workers observability: tail logs, list deploys, run bindings
D1 SQL queries (SELECT and mutations)
R2 and KV CRUD from prompts
DNS and Radar for zone and internet-health queries

Живое демо

Как выглядит на практике

cloudflare.replay ▶ готово

0/0

Установка

Выберите клиент

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "cloudflare": {
      "command": "npx",
      "args": [
        "-y",
        "@cloudflare/mcp-server-cloudflare"
      ]
    }
  }
}

Откройте Claude Desktop → Settings → Developer → Edit Config. Перезапустите после сохранения.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "cloudflare": {
      "command": "npx",
      "args": [
        "-y",
        "@cloudflare/mcp-server-cloudflare"
      ]
    }
  }
}

Cursor использует ту же схему mcpServers, что и Claude Desktop. Конфиг проекта приоритетнее глобального.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "cloudflare": {
      "command": "npx",
      "args": [
        "-y",
        "@cloudflare/mcp-server-cloudflare"
      ]
    }
  }
}

Щёлкните значок MCP Servers на боковой панели Cline, затем "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "cloudflare": {
      "command": "npx",
      "args": [
        "-y",
        "@cloudflare/mcp-server-cloudflare"
      ]
    }
  }
}

Тот же формат, что и Claude Desktop. Перезапустите Windsurf для применения.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "cloudflare",
      "command": "npx",
      "args": [
        "-y",
        "@cloudflare/mcp-server-cloudflare"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json

{
  "context_servers": {
    "cloudflare": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "@cloudflare/mcp-server-cloudflare"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add cloudflare -- npx -y @cloudflare/mcp-server-cloudflare

Однострочная команда. Проверить: claude mcp list. Удалить: claude mcp remove.

Сценарии использования

Реальные сценарии: Cloudflare

Debug a Worker that's throwing 500s in production

👤 Cloudflare Workers developers on call ⏱ ~15 min intermediate

Когда использовать: Your Worker's error rate spiked. You want logs, recent deploys, and a diff of what changed — without opening the dashboard.

Предварительные требования

Cloudflare account OAuth-connected to your MCP client — First tool call triggers OAuth; grant the 'Workers Observability' and 'Workers Bindings' scopes

Поток

Tail recent Worker logs filtered by error

Tail logs for Worker 'api-edge' in the last 15 minutes. Filter to status >= 500. Group by the first 100 chars of the error message.✓ Скопировано

→ Top error templates with counts and timestamps
List recent deployments

List the last 5 deployments of 'api-edge'. Show deploy time, author, and the version hash.✓ Скопировано

→ Deploy timeline — correlate with error onset
Roll back if needed

The error spike starts after the deploy at 14:22. Roll 'api-edge' back to the previous version. Ask me before confirming.✓ Скопировано

→ Confirmation prompt before destructive action

Итог: A restored production Worker, with a clear 'deploy X caused errors Y' postmortem note.

Подводные камни

Log tail is real-time only; can miss a burst that already passed — For historical windows, use the Logpush or Analytics Engine MCP tools instead of tail
Rollback doesn't migrate D1/KV state — If the bad deploy ran migrations, rolling the Worker alone isn't enough — you may need a D1 restore too

Сочетать с: github · sentry

Run ad-hoc analytics against a D1 database

👤 Developers using D1 for app data ⏱ ~10 min beginner

Когда использовать: You want signup-conversion or usage stats from your D1, without writing a dashboard.

Поток

Find the right database and list schema

List my D1 databases. For the one called 'prod-app', show all tables and their columns.✓ Скопировано

→ Database inventory plus schemas
Run the analytics query

In D1 'prod-app', count users who signed up in the last 30 days grouped by week. Show only users who have at least one event in the events table.✓ Скопировано

→ Per-week counts, valid SQL
Iterate

Break it down further by signup source. Which source has the best 7-day activation rate?✓ Скопировано

→ Per-source comparison with rates

Итог: Decision-ready numbers with the SQL shown.

Подводные камни

D1 has per-query row and execution-time limits — For large aggregations, pre-aggregate into a summary table on a schedule instead of scanning raw events each time

Сочетать с: notion

Audit and clean up a bloated KV namespace

👤 Engineers whose Workers cache has drifted ⏱ ~20 min intermediate

Когда использовать: Your KV bill is up; you suspect stale keys or misconfigured TTLs.

Поток

Survey the namespace

For KV namespace 'session-cache', list the first 1000 keys. Sample 10 values and tell me their structure.✓ Скопировано

→ Key pattern distribution, sample value shapes
Identify stale entries

For keys matching session:*, how many haven't been accessed in >30 days? (Use metadata if present; otherwise sample and check timestamps in values.)✓ Скопировано

→ Stale key estimate plus criteria used
Delete safely

Delete keys matching session:expired:* in batches of 100. Show me the first batch before proceeding.✓ Скопировано

→ Batch preview before any deletion

Итог: A cleaner KV namespace with lower storage cost.

Подводные камни

KV is eventually consistent — deletes can appear to un-delete briefly from edge POPs — Don't rely on immediate consistency after bulk delete; verify state after a minute

Review DNS records before a domain migration

👤 Ops engineers migrating domains between providers ⏱ ~15 min beginner

Когда использовать: You're about to change nameservers and you want to inventory every record so you don't drop MX, DMARC, or a forgotten subdomain.

Поток

Dump every record

List all DNS records for zone example.com. Group by type. Include priorities for MX and weights for SRV.✓ Скопировано

→ Complete record inventory
Flag critical records

Highlight any that would break email (MX, SPF in TXT, DKIM, DMARC), plus any A/AAAA/CNAME pointing at third-party services (Stripe, HubSpot, status page).✓ Скопировано

→ Critical-record list with rationale
Produce a migration checklist

Turn this into a checklist I can run on the new provider — each record with its destination, TTL, and 'test after migration' step.✓ Скопировано

→ Copy-paste runbook

Итог: A migration-day runbook that leaves no record behind.

Подводные камни

Forgotten DKIM records break email silently 24h later — Specifically list every _domainkey.<selector> record — they're the easiest to miss

Сочетать с: filesystem

Check Cloudflare Radar for an internet incident affecting users

👤 Incident responders, support leads ⏱ ~10 min intermediate

Когда использовать: Users are reporting your site is down in Brazil. Could be you, could be the internet.

Поток

Query Radar for the country

Cloudflare Radar: is there any notable internet disruption in Brazil in the last 6 hours? Include BGP anomalies, ISP outages, attack traffic.✓ Скопировано

→ Known events list or 'nothing anomalous'
Cross-check with your traffic

For my zone example.com, traffic from BR in the last 6 hours — volume, HTTP status breakdown, top user agents.✓ Скопировано

→ BR-specific traffic profile
Conclude

Based on Radar + my traffic, is this a general Brazil-internet issue or specific to my site?✓ Скопировано

→ Clear verdict with the supporting evidence

Итог: A defensible 'not our fault, it's <ISP X> issue' or 'it IS our fault, here's what' answer.

Подводные камни

Radar data has ~1h lag — For ultra-fresh incidents, pair with your own RUM data

Сочетать с: sentry

Комбинации

Сочетайте с другими MCP — эффект x10

cloudflare + github

Correlate a Workers rollback with the GitHub PR that introduced the regression

Worker 'api-edge' is failing since deploy at 14:22. Find the GitHub PR whose merge corresponds to that deploy and summarize its changes.✓ Скопировано

cloudflare + sentry

Sentry reports Worker-origin errors; Cloudflare MCP pulls the Worker-side logs for the same requestId

Sentry issue EDGE-441 has CF-Ray 8abc123. Tail Worker 'api-edge' for that ray and show the matching log line.✓ Скопировано

cloudflare + filesystem

Edit Worker source locally, deploy via wrangler, verify via Cloudflare MCP logs

Fix the bug in src/index.ts at line 44, deploy, then tail 'api-edge' logs to confirm no more 500s.✓ Скопировано

Инструменты

Что предоставляет этот MCP

Инструмент	Входные данные	Когда вызывать	Стоимость
workers_tail	script_name: str, filter?: object	Real-time Worker log tailing	free (within plan limits)
workers_list_deployments	script_name: str	Review recent versions of a Worker	free
workers_rollback	script_name: str, version_id: str	Revert to a prior version — gated, destructive	free
d1_list_databases	none	Inventory your D1 databases	free
d1_query	database_id: str, sql: str, params?: []	Run SELECT or mutating SQL — writes require explicit prompt confirmation	D1 pricing per rows read/written
kv_list_keys	namespace_id: str, prefix?: str, limit?: int, cursor?: str	Enumerate keys for audit	KV read pricing
kv_get_value / kv_put_value / kv_delete	namespace_id, key, value?, ttl?	Read/write/delete specific keys	KV op pricing
r2_list_buckets / r2_list_objects	bucket?, prefix?	R2 inventory	R2 read ops
r2_get_object / r2_put_object	bucket, key, body?	Read/write R2 objects	R2 op pricing
dns_list_records	zone_id: str	Zone inventory	free
dns_create_record / update_record / delete_record	zone_id, record params	Zone mutations — gated	free
radar_get_http_timeseries / radar_get_attacks	timeframe, region filters	Global internet health context	free
analytics_engine_query	sql: str	Custom Workers Analytics Engine queries	analytics engine read ops

Стоимость и лимиты

Во что обходится

Квота API: Per-plan Workers/D1/KV/R2 limits apply; MCP calls count against your regular usage
Токенов на вызов: 200-2000 tokens typical; log tails can be very large — always filter
Деньги: MCP is free; your Cloudflare services bill as usual
Совет: D1 and KV bill by row-read and op-count. Bulk list/scan can be surprisingly expensive — paginate with modest page sizes and stop early.

Безопасность

Права, секреты, радиус поражения

Минимальные скоупы: Workers:Read D1:Read KV:Read

Хранение учётных данных: OAuth tokens managed by your MCP client; no long-lived API tokens in env

Исходящий трафик: All calls to Cloudflare API; OAuth flow goes through dash.cloudflare.com

Никогда не давайте: Account:Admin unless absolutely needed Zone:DNS:Edit on live zones without a staging test first

Cloudflare MCP is remote/hosted — every tool call is an authenticated request to Cloudflare. Review OAuth scopes carefully at first connect.
Workers rollback is instant and global. Confirm before invoking during peak traffic.

Устранение неполадок

Частые ошибки и исправления

OAuth flow didn't complete

Your MCP client may not support OAuth redirects. Check its docs (Claude Desktop, Cursor, etc. all support it differently). Try re-connecting the remote MCP from the client UI.

Workers tail disconnects after a minute

Tail sessions are time-bounded. Restart the tail, or for longer windows use Logpush and query via Analytics Engine instead.

D1 query returns 'Too many rows read'

D1 caps rows-scanned per query by plan. Add a WHERE clause that uses an index, or paginate with LIMIT.

permission denied on a DNS tool

The OAuth scope for Zone:DNS:Edit wasn't granted. Reconnect the MCP and approve the additional scope.

Альтернативы

Cloudflare в сравнении

Альтернатива	Когда использовать	Компромисс
AWS MCP (awslabs)	You're on AWS, not Cloudflare	Different cloud surface; not a drop-in
Vercel MCP	Your deploy target is Vercel (edge functions, KV, blobs)	Similar remote MCP model; narrower feature set
wrangler CLI directly via shell	You want full wrangler power (config edits, secrets) not just the MCP surface	No agent ergonomics; wider blast radius if scripted wrong

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills