Ponytail for Claude Code: what it is, how to install it, and how much it really saves
Ponytail is a Claude Code plugin that makes the agent write only the code a task needs. Install commands, the lite / full / ultra levels, how to read the author's token-saving benchmark, and how to check the savings on your own bill.
Ponytail is a Claude Code plugin that makes the agent write only the code a task actually needs. At its core is a rule set (skills/ponytail/SKILL.md in the repo): don't write more than necessary, but, as the author stresses, never skip validation, error handling, security or accessibility. Less code means fewer tokens, less time and a smaller bill.
The project is DietrichGebert/ponytail on GitHub, under the MIT licence. Besides Claude Code it supports Codex, Cursor, OpenCode and other tools.
Install (Claude Code)
Send these in Claude Code as two separate messages:
/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail
This adds six skills and two hooks that switch ponytail mode on at the start of each session.
A few things to watch:
- You need Node.js. The hooks run with
node, and it has to be on the non-interactive shell's PATH (nvm and Nix users, take note). Without it the skills still work, but every hook call printsnode: command not found. - The desktop app works too. In the Code tab of the Claude Code desktop app, type the same two commands into the prompt box, or click + next to it and use Plugins → Add plugin.
- Use the original repo. The author asks you to install only from
DietrichGebert/ponytailon GitHub or@dietrichgebert/ponytailon npm. There are several forks with the same name, so check the owner first.
For Codex, run codex plugin marketplace add DietrichGebert/ponytail and codex plugin add ponytail@ponytail, open /hooks in Codex, trust its two hooks and start a new thread.
How to use it
| Command | What it does |
|---|---|
/ponytail lite / full / ultra | Set the level; each step is leaner |
/ponytail off | Turn it off |
/ponytail-review | Review the current diff for bugs, security issues and missing tests |
/ponytail-audit | Run the same review on the whole repo |
/ponytail-debt | Collect shortcut: comments into a debt list |
/ponytail-gain | Show the author's benchmark scoreboard |
/ponytail band on / off | Show cost, tokens and time for the current task above the prompt |
Set a default level in ~/.config/ponytail/config.json or with the PONYTAIL_DEFAULT_MODE environment variable.
The status band needs Claude Code 2.1.290 or later and does not show in the VS Code panel, cloud sessions or claude -p.
Choosing a level: full or ultra suit bug fixes, small features and prototypes. When you need a complete project scaffold, sample code or teaching code that others will read, very lean code can be harder to follow, so use lite or turn it off.
How much it saves: reading the author's numbers
These are the author's own results from the ponytail README (checked 2026-11-07). The method: the same 39 tasks (including a FastAPI + React project), run 5 times each in Claude Code with and without ponytail.
| Metric | Compared with no plugin |
|---|---|
| Lines of code | about 53% fewer |
| Tokens | about 45% fewer |
| Time | about 41% less |
| Cost | about 26% lower |
| Risky logic shipped with a test | 98% (68% without) |
| Bugs caught by the agent's own tests | 66% (46% without) |
Three things to keep in mind:
- These are self-reported and not yet independently reproduced. The full method and per-task results are in the repo's
benchmarks/results/folder. - The widely shared "94% less code" is a single-task ceiling, as the README itself now says. The averages are the ones above.
- Cost falls less than tokens, so not all of the saving is in the expensive part. What you save depends on the kind of task and the level you use.
Check it against your own bill
The most reliable test is your own work:
- Pick a task you do often, such as "add input validation and tests to an endpoint".
- Run it once with
/ponytail offand note the start and end times. - Run a similar task with
/ponytail full. - Compare tokens and cost between the two runs.
If your Claude Code runs on a per-token API, every request's input tokens, output tokens and charge are on your bill, which makes the comparison easy. With CCSub, for example, the Usage page in the dashboard lists tokens and charge for every request; filter each run's time window and add them up. The cost shown in the status band is the plugin's own estimate; your provider's bill is the source of truth.
Uninstall
First run node scripts/uninstall.js to remove the files it writes outside the plugin folder (the mode flag, the config file, the statusline script), then:
/plugin remove ponytail
Keep that order: the cleanup script is itself a plugin file, so removing the plugin first deletes it. You can also run the script from a separate clone of the repo.
Does it depend on your API provider?
No. Ponytail only changes how the agent writes code and does not touch your API settings, so it installs and behaves the same whether Claude Code uses an official account or a third-party service. If you pay per token, what it saves shows up directly on your bill.
If you have not set up Claude Code yet, see the Claude Code setup guide. If you use both Claude Code and Codex, the CC Switch guide helps you manage both configs in one place.
Related setup guides and fixes
Related Posts
Taste Skill for Claude Code: what v2.5 does, how to install it, and where the old version went
Taste Skill is an open-source Claude Code skill that takes the generic AI look out of generated web pages. How the current 2.5 release differs from the version most tutorials show, how to install each, what it needs to run, and how it compares with the official frontend-design plugin.
Using Claude Code from China: 5-Minute Setup with CCSub Relay
Step-by-step guide to using Claude Code in China without VPN. Sign up for CCSub, create an API key, and configure environment variables. Works on macOS, Windows, Linux, and the VS Code extension.
Claude API Proxy: The Complete 2026 Guide (What It Is, Why Use One, How to Set It Up)
A Claude API proxy gives you a single Anthropic-compatible endpoint for Claude Code, Cursor, and Codex — with region access, cost control, and automatic failover. This guide explains what a Claude API proxy is, the self-hosted vs managed options, and a 5-minute setup with CCSub.