ISSUE 067  ·  17 NOVEMBER 2026 ELEN RSS
CodeCity
DEVELOPMENT № 0044

cf: Cloudflare’s agent-first CLI that does not guess its own flags

cf is a new command-line tool from Cloudflare, built for agents: 128 top-level commands, JSON output by default and natural-language command search, so an agent finds the right command instead of guessing it. It is a separate program with its own name, not an extension of Wrangler. In our test, installing with a pinned version took 13 seconds and landed at 169 MB, the full command list came out at 17,160 bytes, and a wrong command answered "Did you mean namespaces?" with exit code 1. The project is five weeks old, with 970 stars and 96 open issues.

cf: το νέο CLI της Cloudflare που δεν μαντεύει flags

What it does

Anyone who has pointed an agent at Cloudflare knows the pattern: you ask for something simple, the agent writes a command from memory, and the command blows up.

I made that mistake by hand, using the syntax Wrangler users have learned, in the singular:

cf kv namespace list

cf did not spit out an unreadable failure. It printed its own list of subcommands, with the word namespaces in the plural, and the line “Did you mean namespaces?”. Exit code 1. The command failed, but the tool told me which one works.

That is the identity of cf: a command-line tool from Cloudflare, built for a user that is not a human but an agent — a program that runs commands on your behalf, powered by a language model. Claude Code, Codex CLI and Cursor are such programs.

Three things serve the decision to stop the agent guessing. First, discovery: cf cli search takes plain language and answers in JSON with a command name and a summary — I asked about “kv namespace” and got the right commands, the correct plural form first, in 488 bytes. Second, structured output: JSON by default, so the agent filters fields instead of guessing columns. There is also a flag that shows the request before it leaves. Third, help in layers: the full list is 17,160 bytes with 128 top-level commands, while one subsystem comes out at 1,470 bytes.

One detail confuses people: cf is not an MCP server, meaning it is not declared in a config file as a source of tools; it is a program you run. The cf mcp command manages Cloudflare Access MCP portals, other people’s servers.

Why a new name instead of extending Wrangler? The vendor frames it as a design decision: Wrangler grew command by command, each product team with its own habits, reaching roughly 280 command paths with uneven terminology. With a fresh name, the team escapes what models have already learned about Wrangler. When a tool changes, the expensive part is the old habits.

On GitHub the repository holds 970 stars, 61 forks and 96 open issues, under an Apache-2.0 licence; it was created on 21 September 2026 and its last push is today, 30 October 2026 (repository). On npm the latest version is 1.0.0-beta.26, published on 29 October 2026, 20:18 UTC, under MIT OR Apache-2.0 (package). The registry holds 49 versions, eight of them inside ten days.

Who it is for

For anyone who touches Cloudflare through an agent or an automation script.

The beginner gains a first step with no account: help, search and local runs all work without keys — in our test, a command with the local simulation flag answered in 1.70 seconds with no credentials at all.

The professional gains output a machine can read, with fewer tokens, plus a way to see the request before it leaves for the API.

It is not for you if you run a mature Wrangler project: its maintenance continues for eighteen months after the beta ends, and a final major version will point users towards cf (announcement). Nor if stability comes first: the version carries the word beta and the project is five weeks old.

Install

Installation goes through npm, from any folder. Prerequisite: Node in one of the ranges ^22.23.0, ^24.22.0 or ^26.12.0. That is a narrow window; outside it npm prints an EBADENGINE warning and carries on, as we saw.

npm install -g cf@1.0.0-beta.26

Why the pin is not a detail

With a bare cf you do not get beta.26 but 0.15.0, from 25 September 2026. The registry’s latest tag points at beta.26, yet bare cf does not follow it: npm picks the newest stable release and skips pre-releases. I tried both routes — bare cf installs 0.15.0, cf@latest installs beta.26.

Pinned: 169 MB, 2,340 files, 31 packages, in 13 seconds. Unpinned: 293 MB, 2,450 files, 34 packages. Almost double the weight, older code. For CI, the chain that builds and tests your code, that means a pipeline pinned to a different version from the one you tested.

The workerd install script

workerd simulates the Workers runtime locally and ships with an install script, code that runs after the package is downloaded. npm 11 does not run such scripts without approval, for security reasons, and prints the line “npm warn install-scripts 1 package has install scripts not yet covered by allowScripts”.

The route npm itself suggests is to grant approval at install time:

npm install -g cf@1.0.0-beta.26 --allow-scripts=workerd

I tested it: the warning goes away, the footprint stays at 169 MB and the package count at 31. A trap not every guide mentions: npm install-scripts approve workerd works in local projects but not for global installs, where it fails with an EGLOBAL error. And the useful part: with the script blocked, the workerd launcher still answered normally with its version. The warning is real, but it does not break the job.

Logging into your account

The program costs nothing; using it needs a Cloudflare account. Without credentials, every meaningful command stops:

cf auth login

In automation with nobody watching, you put a token in the CLOUDFLARE_API_TOKEN variable. The way you define a variable differs from shell to shell, which is why I do not give it as a command to copy. The tool also printed a third route, with no account: a --temporary-account flag with a temporary account and a claim link. We did not follow it to the end, so it stays a promise.

Using it per LLM

cf is not an MCP server: you declare nothing in a config file, you restart nothing. The assistant hands a command to the shell and reads what comes back.

The vendor ships ready-made instructions for the assistants’ rules files, titled “Cloudflare CLI – cf – v20260928”: prefer cf over Wrangler, use cf init, cf dev, cf deploy, and check instead of guessing (announcement). The core move is the same everywhere:

cf cli search "deploy a worker"

Claude Code

In CLAUDE.md you write the convention: for Cloudflare use cf, and when you do not know the syntax, ask first.

cf cli search "deploy a worker"

Codex CLI

The same shape in an AGENTS.md file. The cf repository carries its own AGENTS.md at the root.

cf cli search "deploy a worker"

Cursor

One instruction in the project rules. For a quick look at what exists:

cf --help

In CI, with nobody in front of it

Here the pin and the token work together. The install inside the pipeline:

npm install -g cf@1.0.0-beta.26

And the connection through the CLOUDFLARE_API_TOKEN variable, supplied by the pipeline’s environment. That is my recommendation, not the vendor’s instruction: do not let the pipeline pick a version.

Where it wins, where it loses

It wins on correction: a tool that tells you “did you mean namespaces?” saves rounds of trial and error. It wins on economy: search returns 488 bytes for five results, where the full list is 17,160 bytes and one subsystem 1,470 bytes. It wins on asking nothing before you look — help, search and local runs work without an account — and on licensing, declared MIT OR Apache-2.0 on npm and Apache-2.0 in the repository.

Now where it loses. First, its age: five weeks old, 96 open issues, the word beta in the version. Not a verdict, a risk you price against what you run in production. Second, the pin trap: miss it and you take 293 MB instead of 169 MB. Third, the workerd warning, which frightens people for nothing. Fourth, the account: without a key, work stops with a message rather than a result.

And the honest comparison: cf is the tail. From 1 September to 28 October 2026 it took 1,356,224 downloads against Wrangler’s 191,377,940 — two orders of magnitude. The figure does not separate automated tests from real users, so we hold it as a ceiling.

Testing

I installed the tool and pushed it step by step. The numbers in this section came out of the runs below, not out of documentation.

npm install -g cf@1.0.0-beta.26

The install finished in 13 seconds, with 31 packages and 169 MB on disk, plus two warnings: a Node engine outside the allowed window, and a workerd script without approval.

cf --version

1.0.0-beta.26, in 0.51 seconds.

cf --help

17,160 bytes, 128 top-level commands, in 0.74 seconds.

cf workers --help

1,470 bytes — roughly eleven to one against the full list.

Then the wall, as it should be:

cf workers list

In 1.65 seconds it printed a box titled “No authentication token found” with two options: the CLOUDFLARE_API_TOKEN variable, or logging in with cf auth login. No hang. Login state also reports itself:

cf auth whoami

{“authenticated”: false, “error”: “Not logged in”}, in 0.79 seconds. The first check without an account finishes in under two seconds.

And the part that works without an account:

cf d1 list --local

An empty list in 1.70 seconds, with no credentials.

Finally my mistakes, the most useful part. Out of Wrangler habit I asked for cf kv namespace list: exit code 1, with the list of correct subcommands and the line “Did you mean namespaces?”. Then I asked the tool:

cf cli search "kv namespace"

JSON with five commands, in the correct plural form, in 488 bytes.

What I did not measure: I did not deploy real code to a Cloudflare account, I did not test the migration from Wrangler, and I did not follow the temporary-account flag to the end. I did not measure memory — a figure from estimation would be invention. Any number not stated here is not ours.

What it costs

What I measured.

Disk: 169 MB, 2,340 files, 31 packages pinned at beta.26, and 293 MB, 2,450 files, 34 packages unpinned, with 0.15.0. Install time: 13 seconds.

Run time: 0.51 seconds for the version, 0.74 for full help, 1.06 for a subsystem, 1.65 up to the message asking for a key, 0.79 for the login check, 1.70 for a local command.

Weight in the assistant’s context: 17,160 bytes for the full list, 1,470 bytes for one subsystem, 488 bytes for a search of five results. At a rough four characters per token, the full list is about 4,300 tokens — an estimate, not a measurement.

The model: zero tokens. cf is not an MCP server, it loads no tool schema, it takes no room in any session, and it wants no graphics card.

Node: ^22.23.0, ^24.22.0 or ^26.12.0. Licences: MIT OR Apache-2.0 on npm, Apache-2.0 in the repository. Account: free, but required for meaningful use.

What the vendor says, without my verification.

That cf covers more than 3,000 API operations against Wrangler’s roughly 280 commands. That JSON and type-checked TypeScript configuration cut the agent’s mistakes. That Vite becomes the default local development environment. That migration runs through cf migrate, and that Wrangler maintenance lasts eighteen months after the beta ends.

I did not test the free-tier limits per command. And the download figures — 1,356,224 for cf, 191,377,940 for Wrangler, from 1 September to 28 October 2026 — are a ceiling, because they do not separate automated tests from mirrors.

Verdict

Score: 7.2/10.

The idea is right and it states itself clearly: a tool the agent can interrogate instead of guess at. In our test that showed twice — the wrong command answered “Did you mean namespaces?” instead of blowing up, and search returned the correct syntax in 488 bytes. Output is JSON, help is layered, and the first check without an account finishes in under two seconds.

What it lacks is years. Five weeks old, 96 open issues, the word beta in the version, and 1.36 million downloads against Wrangler’s 191.4 million.

In practice: try it on new work, not in production. Pin 1.0.0-beta.26, because a bare cf gets you 293 MB instead of 169 MB. Ask with cf cli search instead of guessing, look at the request before any command that writes, and keep Wrangler alongside for whatever is already in production.

Sources: repository https://github.com/cloudflare/cf · npm package https://www.npmjs.com/package/cf · Cloudflare announcement https://blog.cloudflare.com/cloudflare-cf-cli-launch · InfoQ https://www.infoq.com/news/2026/10/cloudflare-cf-cli/ · npm downloads https://api.npmjs.org/downloads/point/2026-09-01:2026-10-28/cf