feiskyer/claude-code-settings: A Plugin Bundle for Claude Code Skills, Agents and Model Routing

What the plugin bundle actually adds to a Claude Code session

Claude Code on its own is a coding agent with a fixed toolset. This repository is a distribution layer on top of it: thirteen skills, a set of sub-agents, guidance documents, and a settings template. The README frames skills as "reusable capability modules" that teach Claude how to perform a specific task, invoked either manually as /skill-name [arguments] or triggered automatically from context.

The useful part is that the skills are not all coding tasks. deep-research splits a research goal into parallel sub-goals and runs them through claude -p subprocesses, aggregating results into a final report. youtube-transcribe-skill pulls subtitles from a YouTube link, preferring a yt-dlp CLI path and falling back to browser automation. Two image skills cover Gemini and OpenAI image models. translate converts English or Japanese technical articles into Chinese using a three-step process that keeps Markdown structure and technical terms in their original form. github-fix-issue and github-review-pr handle repository automation.

The audience is narrow and specific: people already using Claude Code who want a curated set of these capabilities rather than assembling them one at a time. If you are evaluating Claude Code itself, this repository assumes you have already made that decision.

How skills, sub-agents and the gateway fit together

The repository layout tells most of the story. skills/ holds the individual capability modules, agents/ holds sub-agent definitions, plugins/ and .claude-plugin/ hold the marketplace packaging, guidances/ holds supporting documentation, and settings.json at the root is the author's configuration template. .mcp.json configures MCP servers, which matters because deep-research documents a tool-selection order: already-installed skills first, then MCP tools, then WebFetch and WebSearch.

The model routing layer is the part worth understanding before you install. The template settings.json points ANTHROPIC_BASE_URL at a copilot-gateway proxy on http://localhost:4141, and the README says you can substitute another gateway such as a LiteLLM Proxy Server by changing that value, giving http://localhost:4000 as the example. Model selection is expressed through three variables: ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL and ANTHROPIC_DEFAULT_HAIKU_MODEL, listed in the README as claude-sonnet-5, claude-opus-5 and claude-haiku-4-5.

So the architecture is: Claude Code talks to a local proxy instead of Anthropic directly, the proxy decides which upstream model serves each tier, and the skills layer on top calls tools that may themselves shell out to other CLIs. Each layer can fail independently, which is worth remembering when something does not work.

Installing through the plugin marketplace and setting up settings.json

The README gives two equivalent marketplace paths. From inside a Claude Code session you use slash commands:

/plugin marketplace add feiskyer/claude-code-settings
/plugin install claude-code-settings

From a shell instead, the same operations run through the CLI. The install command here names the plugin with its marketplace suffix:

claude plugin marketplace add feiskyer/claude-code-settings
claude plugin install claude-code-settings@claude-code-settings

After either path, the main plugin brings in the full skill list. The README is explicit that this does not cover everything: settings.json is not configured through the plugin and must be set manually. That is the step people miss. Copy the template, then edit the model names and the base URL to match your own gateway.

There is a third route for installing individual skills, via npx skills, and the README warns that versions installed this way may lag behind the repository. The manual installation route is the most invasive: it moves ~/.claude to ~/.claude.bak and clones the repository into ~/.claude, then starts the gateway with npx copilot-gateway@latest start --proxy-env. First launch prompts for GitHub device authentication. Treat that route as a last resort, because it replaces your entire configuration directory.

Where the defaults can bite you: models, permissions and the manual settings step

Three limitations stand out.

First, the model names are assumptions. The README states plainly that you must ensure the listed models are available in your account, otherwise replace them with your own model names as actually provided by your gateway. If you point ANTHROPIC_BASE_URL at a LiteLLM proxy that does not serve claude-sonnet-5, the template will not work as written. This is not a bug; it is a template doing what templates do. But it means the install is not finished when the plugin install returns.

Second, the permission mode. The template sets defaultMode to acceptEdits, which the README describes as automatically accepting file edits while keeping command confirmation. It notes you can change this to bypassPermissions to skip confirmation entirely, and asks you to understand the security risk first. That is the correct framing, and it is also a genuine decision point: acceptEdits means edits land without review while shell commands still prompt.

Third, the settings file is outside the plugin's lifecycle. Plugin install and update will not touch it, so upgrading the plugin does not upgrade your configuration, and a stale settings.json can silently keep pointing at a gateway you stopped running. The README does not document a rollback procedure for the manual installation path beyond the initial backup move.

One more constraint worth noting: several skills have their own dependencies that the plugin does not install. codex-skill needs the Codex CLI, nanobanana-skill needs GEMINI_API_KEY in ~/.nanobanana.env plus Python packages, gpt-image-skill needs OPENAI_API_KEY, and youtube-transcribe-skill needs yt-dlp or chrome-devtools-mcp. A plugin install gets you the instructions, not the runtime.

npx skills versus the marketplace plugin: two distribution models

The most direct alternative to this repository is not another project but the npx skills installer that the README itself documents as an alternate path. The difference is granularity and freshness. The marketplace plugin installs the whole bundle as one unit and tracks the repository version. npx skills installs individual skills on demand, with subcommands for listing, installing everything, or interactively choosing, but the README warns that the versions it serves may lag behind the repository.

That trade-off is real. If you want one skill, say youtube-transcribe-skill, pulling the entire plugin bundle means carrying skills you will never invoke, along with their trigger words. If you want the current version of a skill, the marketplace path is the one the README recommends. There is no option that gives you both per-skill granularity and current versions.

A second comparison point is the sibling repository. The README points to feiskyer/codex-settings for OpenAI Codex configuration and custom prompts. If your agent of choice is Codex rather than Claude Code, that is the repository to look at; the skills here are written for Claude Code's skill format and slash-command invocation, and do not transfer.

Maintenance and licence

The repository is not archived, and the last push was on 2026-08-13.

The licence is MIT, stated in the README badge and present as a LICENSE file at the repository root. MIT permits use, modification and redistribution with the copyright notice and permission notice retained. That matters here because the repository contains a settings template and skill definitions you are expected to copy and edit into your own configuration; a more restrictive licence would complicate that workflow. This is a description of the licence terms, not legal advice, and the skills invoke third-party services under their own terms: Google Gemini, OpenAI, GitHub Copilot and YouTube each have their own conditions that the MIT licence does not cover.

Upgrade cost is low for the plugin and non-trivial for the settings file. Plugin updates flow through the marketplace commands. settings.json does not, so any change to model names, base URL or permission mode is yours to make and yours to maintain across upgrades. There have been no retrieved releases, so the plugin appears to be distributed from the repository and marketplace rather than through tagged versions.

Editorial conclusion

Adopt it if you already run Claude Code and want deep-research, GitHub issue and PR automation, image generation or translation available as slash commands without writing each skill yourself. Skip it if you do not want a third-party settings.json template deciding your model names and permission mode, or if your gateway does not serve the model identifiers the template lists. Before installing, verify three things: that the models named in settings.json exist on your account or gateway, that you have backed up ~/.claude because manual installation overwrites it, and that you are comfortable with defaultMode set to acceptEdits rather than bypassPermissions. The plugin itself installs from the marketplace; the settings.json file does not.

Frequently asked questions

Official sources

1. feiskyer/claude-code-settings on GitHub
2. Issues
3. License: MIT
4. Project website
5. README