AI Guides › Workbench
By Nigel Guy · 7 min read
The usual way into add-ons is a list post: four names, four install lines, paste them all in one go and carry on vibe coding. Then Claude starts behaving differently, your usage drains faster or slower and you cannot tell which add-on did what. Worse, one of the four quietly changes where your requests are sent.
The rule: install one add-on at a time, read what it will run before you say yes, and keep the one that changes where your code goes switched off unless you have a reason to use it.
Every command below was checked against each project's README or the Claude Code docs on 4 October 2026. Re-read the README before you run anything; these projects change weekly.
| Word | Plain meaning |
|---|---|
| Plugin | A package for Claude Code that can add commands, skills, hooks and MCP servers in one install |
| Marketplace | A catalogue of plugins, usually a GitHub repo. You add it once, then install from it |
| Skill | A folder with a SKILL.md file of instructions. Claude loads it when your request matches its description, or when you type /skill-name |
| Hook | A small script Claude Code runs automatically at set moments, such as before a tool call |
| Gateway | A program that sits between Claude Code and the model, and decides where each request goes |
| Add-on | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Ponytail | Skill and hooks that push Claude to reuse what exists and write the smallest change | Free (MIT) | Builds that keep sprouting helper files and new dependencies | Adds two Node hooks; "minimal" can mean skipping error handling you wanted |
| OmniRoute | Local gateway that routes requests to many providers, with automatic fallback when one runs out | Free (MIT); providers' own terms and limits apply | Experimenting with free or cheap models outside Claude | Anthropic does not support Claude Code on non-Claude models; routing a subscription account through it risks your account |
| Graphify | Builds a knowledge graph of your folder so Claude can query it instead of opening files one by one | Free (MIT and Apache 2.0); docs and PDFs use your model | Large or unfamiliar codebases | Writes to CLAUDE.md and adds a hook; the graph goes stale unless you update it |
| Anthropic's example skills | A demonstration set of skills (design, development, communications, documents) that load when your request matches | Free | Seeing what good skills look like | Anthropic labels them for demonstration and education; every skill adds a line to Claude's context |
Inside a Claude Code session, send these as two separate messages. The README says the install fails if you send both at once.
/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail
The install opens a details pane. Read the Will install list, then choose Install for you, in this repo only (local scope) while you trial it. Afterwards /ponytail lite, full, ultra or off sets the strength, and /ponytail-review checks a diff for over-building. The project publishes its own token-saving benchmark; I have not reproduced it, so treat it as the author's claim.
npm install -g omniroute
omniroute setup
omniroute setup is the first-run wizard; the dashboard then runs at http://localhost:20128. omniroute configure claude writes a Claude Code profile, and omniroute doctor checks connectivity.
Read this first. Claude Code's gateway docs say Anthropic "doesn't support routing Claude Code to non-Claude models through any gateway", so fallbacks to other models may break features without warning. OmniRoute's own README warns that routing Claude or ChatGPT subscription accounts through a third-party gateway risks account bans, and flags some free providers as terms risks. Free providers may also log what you send. Use it in a throwaway project, never on client code, and remove the ANTHROPIC_BASE_URL setting when you are done.
In your normal terminal, not inside Claude:
uv tool install graphifyy
graphify install
The package name has two y's; the command has one. (pipx install graphifyy also works.) In Claude Code, type /graphify . to build graphify-out/ with graph.html, GRAPH_REPORT.md and graph.json. Optional: graphify claude install adds a CLAUDE.md section and a PreToolUse hook that nudges Claude to the graph before it searches. Code is parsed locally; documents, PDFs and images go to your model. Add a .graphifyignore for secrets and generated folders, run graphify update ./src after big changes, and graphify uninstall removes it.
Fill in [PROJECT_FOLDER] and [QUESTION]:
You are working in [PROJECT_FOLDER], which has a Graphify knowledge graph in graphify-out/.
Goal: answer [QUESTION] using the graph first, opening source files only when the graph cannot settle it.
Steps:
1. Read graphify-out/GRAPH_REPORT.md and query graph.json for the parts that bear on the question.
2. List the files the graph points to, and why each matters, before opening any of them.
3. Open only those files, and say which ones you opened.
Output: a short answer, then a table of files consulted (file, reason, opened yes/no).
Rules: do not edit anything. If graphify-out/ is missing or older than the latest commit, stop and tell me to rebuild rather than guessing.
Before answering, check that every file you name exists and that the answer would hold if the graph were out of date.
The source idea is a bundle that "switches on" skills per stage. In fact that is how Claude Code already treats every skill: it reads each skill's description and loads the one that matches your request. The bundle just gives you more skills to match against.
/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills
Skills from a plugin appear as /plugin-name:skill-name. To see which ones fire at each stage, fill in [STAGE] and [TASK]:
I am at the [STAGE] stage of a small build (plan, build, test or ship). The task is: [TASK].
Before doing anything, list the skills you have available that match this stage, quoting the line of each description that matched, and say which one you would load and why.
If none fits, say so plainly instead of stretching one to fit.
If [STAGE] or [TASK] is unclear, ask me one question first.
Do not start the task until I reply "go".
claude plugin list in your shell shows what is installed after each step. Never install all four in one sitting: you will not know which one changed Claude's behaviour.
/usage instead./reload-plugins, then check the Errors tab in /plugin.claude plugin disable <name>@<marketplace>.ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN and restart Claude Code.