Cloudflare Developer Platform

The Workers Bindings server. Cloudflare publishes seventeen managed MCP servers on separate hostnames, and this listing is the one at bindings.mcp.cloudflare.com. Its siblings cover documentation, observability, Radar, browser rendering, audit logs and more, each with a different tool set. Connecting this one gives you storage and compute primitives only.

It can delete one, and Cloudflare publishes no way to get it back. The kv_namespace_delete tool is annotated destructive in the server's own source, and Cloudflare's Workers KV documentation describes no trash, undelete or restore path for a namespace. There is no inverse tool on this server either.

No. Time Travel restores an existing database to an earlier minute within the last thirty days on Workers Paid, or seven on Workers Free. Every command names a database that must still exist, so it recovers bad writes rather than dropped databases. Cloudflare recommends exporting to R2 for anything longer.

Yes. The d1_database_query tool takes a free-text sql parameter and passes it to Cloudflare's D1 query API unchanged. Nothing on this server restricts statements to reads, so DROP TABLE and DELETE reach production the same way SELECT does. The tool is not annotated destructive in Cloudflare's source.

Automatically, per request. Cloudflare removed the set_active_account and accounts_list tools in workers-bindings 0.5.0. An AccountManager now resolves the account from your token when it covers one account, then a cf-account-id header you set in client config, then an optional account_id argument. No selection persists between calls.

Account read, Workers write and D1 write, plus user read and refresh tokens. Cloudflare defines that set in the server's published source and shows it on the authorization page, where you also scope the grant to one account. The OAuth descriptor omits scopes_supported entirely, so no client can discover them before you consent.

Yes, indirectly. D1 bills rows read and written, and Cloudflare states dashboard and CLI queries count as billable usage, so agent queries do too. Creating R2 buckets and D1 databases adds billable storage. R2 also requires a subscription added through checkout before any bucket exists.

No. Anthropic's directory description offers deploying code to production as an example prompt, but no deploy tool exists in its own tool list or in Cloudflare's published reference. The three Workers tools all read: they list Workers, fetch details and fetch source code. Deployment stays with Wrangler or the dashboard.