Use cf with coding agents

Coding agents can use cf to work with your Cloudflare account. Commands return JSON and describe their own inputs, so an agent can find and run the right command without prior knowledge of cf.

Set up `cf` for your agent

Install cf globally so every agent session can run it:
npm install --global cf
yarn global add cf
pnpm add --global cf
bun add --global cf
Inside a project that installs cf as a dependency, the global command runs the project's copy.
Sign in:
```
cf auth login
```
Sign-in opens a browser page where you approve access, so complete this step yourself. For agents that run without a person present, refer to Authenticate unattended agents.
Tell your agent to prefer cf. Add this instruction to your user-level AGENTS.md, CLAUDE.md, or equivalent instructions file:
AGENTS.mdmd
```
When interacting with Cloudflare, use the `cf` CLI unless the project has a
Wrangler configuration file.
```
Projects that still use wrangler.jsonc or wrangler.toml keep working with Wrangler, while every other Cloudflare task goes through cf.

You can also ask your agent to do this setup for you:

Install the Cloudflare CLI with `npm install -g cf`. Then add this line to my
user-level AGENTS.md or CLAUDE.md file: "When interacting with Cloudflare, use
the cf CLI unless the project has a Wrangler configuration file."

Help your agent find commands

cf has more than 2,900 commands. Instead of reading help for every product, an agent can describe the task, inspect the best match, and then run it.

Search for a command by describing the task:
```
cf cli search "create D1 database"
```
The search runs locally and needs no credentials. Put the task in quotes. cf cli search takes the whole task as one argument, and unquoted words after the first fail as unknown commands. It returns up to five matches as a JSON array, best match first. Each match has the command and a short summary. These are the first two matches for this search:
```
[
  {
    "command": "cf d1 create",
    "summary": "Create D1 Database"
  },
  {
    "command": "cf d1 update",
    "summary": "Update D1 Database"
  }
]
```
Inspect the API request of the chosen command:
```
cf schema d1 create
```
The JSON result describes the API request that the command sends: operationId, httpMethod, path, pathParams, queryParams, hasRequestBody, and requestBodyFields. Each parameter and body field lists its name, its type, and whether it is required. Body fields also include a description. cf schema covers generated API commands. For a command's arguments and options, or for a command such as cf deploy, run the command with --help.

Path and query parameters use their API names, such as account_id. requestBodyFields lists the body fields that the command also accepts as options, under the option names. A field inside an object gets a combined name, such as read-replication-mode for read_replication.mode, and array fields show the type string. For some bodies, the list is incomplete or empty. For these commands, pass the complete body with --body, and check the API reference for its shape.
Preview the request with --dry-run, then run the command without it:
```
cf d1 create --name my-database --dry-run
```

Root and group --help output starts with a reminder for agents to use cf cli search first. A command's own --help shows its usage as plain text. If an agent types a command that does not exist, cf lists the closest matches.

Read command output

Results go to standard output as JSON. Messages, such as the selected zone, and errors go to standard error. When standard output is not a terminal, it contains only the result, so an agent can parse it or filter it with jq without extra flags.

Lists print a JSON array and return one page. Paging options differ between commands, so check the command's --help output.
Changes that return no data print nothing to standard output.
Raw content, such as an R2 object or an image from a Workers AI model, goes to standard output unchanged. Redirect it to a file.
JSON output is indented, whether or not standard output is a terminal. Colors are added only in a terminal.

A failed command exits with a non-zero status and prints the error to standard error.

Run commands safely

Preview changes. Add --dry-run to print the request as JSON without sending it. Dry runs need no credentials.
Check for aborted deletes. In a non-interactive session, a destructive command without --force prints Aborted. to standard error and exits with status 0. A successful exit does not mean the resource was deleted.
Review --force before you allow it. On some commands, --force is also an API parameter. For example, cf workers delete --force also deletes a Worker that other Workers still reference.
Use local data where supported. --local works only for a few resources that local development provides, mainly KV keys, D1 databases through cf d1 raw, cf d1 migrations list, and cf d1 migrations apply, and R2 objects. Commands without a local equivalent, including cf d1 query, return an error instead of calling the Cloudflare API.

For local state, refer to Local resource data.

Work in Wrangler projects

To migrate a project with an agent, ask it to:

Preview the migration without writing files:
```
cf migrate --dry-run
```
Run the migration:
```
cf migrate
```
Resolve the TODO(@cloudflare) comments in the generated cloudflare.config.ts, and ask you about any choice it cannot make on its own. The build fails until every required item is resolved.

For the full process, refer to Migrate a Wrangler project.

Authenticate unattended agents

Agents that run without a person present, such as in continuous integration (CI), cannot complete cf auth login. Set CLOUDFLARE_API_TOKEN instead. The token takes priority over any stored login. For the variables, token permissions, and non-interactive behavior, refer to Use cf in CI.

To give an agent separate credentials for one project on your own machine, bind a named profile to the project directory. Refer to Use named profiles.

Inspect local resources

When cf dev runs inside a supported coding agent, the development server also prints the URL and main routes of the Local Explorer API. The agent can use this API to read and change the data in your Worker's local bindings, and to query local traces and logs.

Agent setup configures Cloudflare Skills and MCP servers for popular coding agents.
Use cf in CI covers API tokens and non-interactive behavior.