Agent Cloudflare

agent-cloudflare

Use agent-cloudflare for Cloudflare operations triage. Prefer investigation commands for incident-shaped questions and direct resource commands when the user names a specific Cloudflare resource.

Safety

• Never ask the tool to reveal an API token.
• Never accept pasted Cloudflare tokens in chat. Ask the user to run agent-cloudflare profiles add <profile> --form locally so the token goes directly into an OS dialog.
• Use agent-cloudflare profiles update <profile> --form when a stored token needs replacement.
• Prefer read-only commands.
• Use --account-id and --zone-id to scope commands when multiple accounts or zones are visible.
• Treat mutations such as DNS changes or cache purges as high stakes. Use --dry-run first and only use --confirm when the user explicitly asks for the write.

Choose The Path

• Setup or credential issue: start with usage, profiles list, profiles check, and profiles discover.
• Incident, outage, or vague question: start with investigate usage, then choose the closest investigation.
• Named resource or known ID: use the matching resource command from references/commands.md.
• Write request: use the dry-run command first, report the planned operation, then wait for explicit approval before --confirm.
• Unsupported Cloudflare endpoint: use api get for authenticated reads and keep output scoped.

First Commands

agent-cloudflare usage
agent-cloudflare profiles list
agent-cloudflare profiles check
agent-cloudflare investigate usage

For local testing, run mockcloudflare and set --base-url http://127.0.0.1:12112 with --api-token cfut_mock.

Output Contract

Lists default to NDJSON. Get (single + multi). get <id>... accepts one or more ids and returns one result per id, in input order. Default output is NDJSON: one line per id — the record, or {"@unresolved":{"id","reason","fixable_by","hint"?}} for an id that couldn't be resolved (e.g. not found). --format json|yaml collapses to one {"data":[…],"@unresolved":[…]} envelope. A single get <id> is the one-element case (NDJSON by default; pass --format json for the object). Item-level misses stay on stdout and exit 0; only command-level failures (auth, network) go to stderr with exit 1. zone-settings get and waiting-rooms get scope their zone via --zone <zone-name-or-id> (not a trailing positional). api get stays single. Errors include fixable_by and usually a hint.

Investigation output uses evidence records:

{"type":"entity","object":"zone","id":"...","data":{}}
{"type":"finding","severity":"warning","summary":"...","data":{}}

Profile/config metadata lives in XDG config. API tokens live in Keychain. profiles list and profiles check may show non-secret credential type (cfut, legacy_api_token, or unknown) but never the token.

Incremental References

Load these only when useful:

• references/scenarios.md: common incident questions and recommended command sequences.
• references/commands.md: command chooser and exact syntax for resources, investigations, raw API reads, and guarded mutations.
• references/output.md: NDJSON/JSON conventions, error hints, evidence rows, secret handling, and mutation approval rules.