Cloudflare Platform

Workers

• V8 isolates, not Node: process, Buffer, fs do not exist unless compatibility_flags = ["nodejs_compat"] is set. compatibility_date must be set or deploys fail.
• Worker size limit is 3 MB compressed on the Free plan, 10 MB on Paid. Keep dependencies minimal.
• ctx.waitUntil(promise) for side effects that must outlive the response (logging, analytics).
• Bindings are reachable only via the handler's env argument, and must exist in both wrangler.toml and the Env interface.
• Durable Objects need [[durable_objects.bindings]] and a [[migrations]] block with new_classes = ["Counter"] — the binding alone will not deploy.

Storage selection

KV: config/sessions/cache, eventually consistent. D1: relational SQLite, strongly consistent. R2: files/blobs, S3-compatible, no egress cost. Durable Objects: coordination, counters, locks, realtime. Queues: async jobs.

Limits: KV value 25MB, key 512 bytes, metadata 1024 bytes, read lag up to 60s. R2 needs multipart upload above 100MB. D1 is SQLite — no MySQL/Postgres-only syntax; always .bind() params; env.DB.batch([...]) collapses round trips.

Wrangler

• wrangler secret put NAME [--env staging] — secrets never live in wrangler.toml; local dev uses .dev.vars.
• KV bindings need preview_id for wrangler dev --remote.
• D1 migrations: pass the target explicitly — wrangler d1 migrations apply <db> --local for dev, --remote for production.
• wrangler tail for live production logs; wrangler deployments list + wrangler rollback <version-id>; wrangler whoami to confirm the account.

MCP

Code Mode exposes exactly two tools: search (Cloudflare API spec) and execute (JavaScript against the API). ~1k tokens per operation vs ~244k for native tool exposure.

Docs: https://developers.cloudflare.com/