Using the tools
How operations are named
Each of the 3,777 operation tools is named after the official Cloudflare SDKs: the resource path, then the method.
| Tool | Operation |
|---|---|
dns_records_list | GET /zones/{zone_id}/dns_records |
dns_records_create | POST /zones/{zone_id}/dns_records |
zones_settings_edit | PATCH /zones/{zone_id}/settings/{setting_id} |
tunnels_config_update | PUT /accounts/{account_id}/cfd_tunnel/{tunnel_id}/configurations |
rulesets_zone_entrypoint_update | PUT /zones/{zone_id}/rulesets/phases/{ruleset_phase}/entrypoint |
The SDKs give the account and zone variants of some resources the same base name. The description says which one a tool is: [zone], [account] or [user]. rulesets_entrypoint_get works on an account, rulesets_zone_entrypoint_get on a zone. When two operations would still collide, one takes its OpenAPI operationId; a name over 60 characters is cut and gets a short hash.
Arguments
zone_idtakes the zone's name (example.com) or its 32-character id.tunnel_idtakes the tunnel's name or its UUID.account_idis optional: without it the server usesCLOUDFLARE_ACCOUNT_ID, or the only account the token sees. If the token sees several, the error lists them with their names.- Body fields are top-level arguments next to the path and query parameters. When a body field would collide with a parameter, or the body is not an object, it goes inside
body. - Query filters that are objects become dotted keys:
name: {"exact": "app.example.com"}is sent asname.exact=app.example.com. - Lists paginate with
pageandper_page; the response carriesresult_info.
A tool returns Cloudflare's result and, when present, result_info. On failure it returns Cloudflare's error codes and messages, with a hint when the status is 401 or 403.
find and call
find searches every operation for a task and returns up to 15 (default 5) with their name, method, path, description and argument schema, plus tips when the task maps to a mechanism worth explaining. call runs the one you pick:
find { "query": "Email Routing catch-all forward" }
→ { "search": "hybrid",
"operations": [ { "name": "email_routing_rules_catch_all_update", … } ],
"tips": [ "Email Routing: the destination must be verified …" ] }
call { "name": "email_routing_rules_catch_all_update",
"arguments": { "zone_id": "example.com", "actions": [ … ], "matchers": [ … ], "enabled": true } }Write the query in English with Cloudflare's product terms, even when the user asked in another language: "WAF custom rules zone entrypoint" finds the right operation far more often than "bloquear um país". When the query looks Portuguese, the result starts with a tip saying so. How find works has the numbers.
call accepts the name with or without a gateway prefix (cloudflare_dns_records_list works too).
Tunnels
tunnel_route_add publishes a hostname on a remotely managed tunnel in one call:
tunnel_route_add { "tunnel": "home", "hostname": "app.example.com", "service": "http://app:8080" }
→ rule added; CNAME created → 9ed7….cfargotunnel.comIt reads the tunnel's configuration, puts the rule before the catch-all (or replaces the rule with the same hostname and path), writes the configuration back keeping everything else, and with create_dns (default true) creates a proxied CNAME in the zone that covers the hostname. It refuses when the last rule is not a catch-all. If another DNS record already holds the hostname, the rule is still written and the result says the DNS step failed and why. tunnel_route_remove removes the rule but leaves the DNS record.
A starter tool set
Showing 3,782 tools to a model costs context on every turn and makes the choice harder. Expose the ones you use daily plus find and call; the rest stays reachable. A set of about 55 that covers most day-to-day work:
find call tunnel_route_add tunnel_route_remove api_request
accounts_list zones_list zones_get zones_settings_*
dns_records_* tunnels_*
zero_trust_access_applications_{list,get,create,update,delete} zero_trust_access_policies_*
cache_purge rulesets_zone_{list,get} rulesets_zone_entrypoint_{get,update} workers_routes_*How you apply it depends on the client: an allow-list of tool names, or an MCP gateway that filters by pattern (for example rules like include dns_records_* followed by exclude *).
Destructive operations
Every DELETE operation, call, api_request and tunnel_route_remove carry destructiveHint: true; GET operations carry readOnlyHint: true. Clients that honour the hints ask before running destructive tools. The token is still the real limit: give it only the permissions you are willing to have used.