Install an open-source Claude Code plugin when it fixes a job you already repeat, from a source you can read, after you have checked what it adds. Anthropic's official marketplace is the safest place to start: Claude Code adds it for you, and Anthropic curates what is listed there. Two of the open-source picks below come from Tigerless Labs, autoharness and auto-gtm. We make them, so we say so in their own section and name the cases where another tool fits better.
Are Claude Code Skills worth using?
Yes, if a skill covers a task you repeat. The words get mixed up, so start with Anthropic's definitions. A plugin is the package you install, and it can hold skills, agents, hooks and MCP servers. A skill is one of those parts: instructions that Claude reads for a kind of task. An MCP server connects Claude to an outside tool or service. The official Discover page puts the relationship in one sentence: plugins extend Claude Code with skills, agents, hooks, and MCP servers (Discover and install prebuilt plugins).
That distinction decides what to try first. A single skill is the cheapest test, because you can read the whole thing in a few minutes before you turn it on. A plugin earns its extra weight when you need several parts to work together, or when you want updates from the marketplace it came from. Start with the smallest thing that solves the job, and move up only when the small thing stops being enough.
Which skills are you using frequently?
We opened the repository or marketplace page for each plugin below. The table shows what each one is for, where it installs from and the license its own page states.
| Plugin | Installs from | License, as its page states it | What it adds |
|---|---|---|---|
| commit-commands | Anthropic's official marketplace, claude-plugins-official | Not checked here; each plugin links its own license | Commands for committing, pushing and opening pull requests |
| Superpowers (obra/superpowers) | GitHub, with its own plugin marketplace | MIT | A development method built on composable skills, plus instructions that make the agent use them |
| Context7 (upstash/context7) | GitHub; an MCP server, not a plugin | MIT | Current library documentation fed to the agent |
| autoharness | GitHub, tigerless-labs/autoharness (ours) | MIT | Skills it learns from your own sessions |
| auto-gtm | GitHub, tigerless-labs/auto-gtm (ours) | MIT | Drafts of X and Reddit posts and replies, built from your pull requests |
Commit-commands is the one to try first from the official set, because Anthropic's own install guide uses it as the worked example (Install and manage plugins). It is small, it does one job, and you can see exactly which commands it adds before you accept it. Superpowers is a larger change: it changes how the agent plans and tests its work, so read its README before you decide you want that. Context7 solves a different problem. It keeps the agent's library knowledge current, but it arrives as an MCP server, so install it with the MCP setup it documents rather than through the plugin steps below.
The official marketplace is curated by Anthropic, and inclusion is at its discretion. The community marketplace, anthropics/claude-plugins-community, hosts third-party plugins that have passed Anthropic's automated validation and safety screening, and each is pinned to a commit (Discover and install prebuilt plugins). Plugins in those two lists have been through a review that a plugin from an arbitrary repository has not, so start there when you can.
How do you install a plugin without trusting it blindly?
Anthropic's documentation is direct about the risk. Plugins and marketplaces can execute arbitrary code on your machine with your user privileges, and Anthropic does not control or verify what a third-party plugin includes (Discover and install prebuilt plugins). Read before you install, then follow these steps in a Claude Code session:
- Open the plugin's repository and read its README, its LICENSE and the hooks it runs.
- Add its marketplace with /plugin marketplace add owner/repo, using the source the README gives.
- Install it with /plugin install plugin-name@marketplace-name.
- In the details pane that opens, read the Will install list: commands, agents, skills, hooks, and MCP and LSP servers.
- Choose a scope, then run /reload-plugins or restart Claude Code so the plugin loads.
Autoharness shows the whole sequence in two lines. It needs Python 3.11 or later as python3 on your PATH, because its hooks do not fire on an older interpreter:
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharnessThen run /reload-plugins. The official install guide describes the same scopes: user for all your projects, project for everyone who works in the repository, and local for one repository only (Install and manage plugins). Pick local while you are still judging a plugin, so the rest of your work is untouched.
How to reduce Claude Code token usage?
Enable fewer plugins. Each one adds to the context Claude Code sends, and the official Discover page shows a context cost estimate for each plugin, described as how many tokens the plugin will add to your context window every turn (Discover and install prebuilt plugins). That cost is paid on every turn, not only on the turns that use the plugin, so a plugin you installed for one task can make every later task more expensive.
Check the estimate before you accept an install, and remove the plugins you stop using. The estimate tells you what a plugin adds to the context; it does not tell you what your bill will be, because that depends on the model, the cache and how long your sessions run. To see what each part of a request costs, read our walk-through of what Claude Code sends to the API. It explains how to capture the request and price the system prompt, tool schemas and MCP servers one by one. We have not measured how much any single plugin saves or costs, and we would not give you a figure without a run behind it.
Which of these do we make?
We make two of the plugins in the table, both under the MIT license and both installed from our own GitHub repositories rather than from Anthropic's marketplace. So the trust you place in them is trust in Tigerless, and you should read them the way you would read any repository you did not write.
Autoharness is a self-learning skill layer for Claude Code. It watches your sessions, distills skills from the work you already do, merges skills that cover the same scenario, and archives the ones that stop being used. It touches only the skills it created, so anything you wrote or installed stays as it is. It has no daemon and no timer, and its README says it is validated by whether the skills are followed in later work rather than by a benchmark. You can read the autoharness page on Tigerless Labs or the repository.
Auto-gtm turns your recent GitHub pull requests and the day's hot threads into topics, then drafts X and Reddit posts and replies from them. It installs in Claude Code and in Codex. Reddit reads go through rdt-cli, which reuses your logged-in browser session, so you do not create an OAuth app or hand over a password. It never posts on its own, and publishing is always manual. See the auto-gtm page or the repository.
Our other Labs projects are not plugins, so they do not belong in this list. agent-memory is a memory store that Claude Code and Codex share through a hook and a command-line tool, seo-ops is a deterministic SEO checker you can run from a skill, and cost-xray is a capture tool that observes spending without changing it. The full list is on the Tigerless Labs page.
When is a plugin the wrong tool?
A plugin is the wrong tool when a file you already write does the job, when the task happens only once, or when you cannot read what the plugin runs. Three cases come up most.
Standing rules belong in files. Build commands, naming conventions and your preferences are better in a CLAUDE.md file or in Claude Code's built-in auto memory than in a plugin. Those two cover most single-person setups. How to give Claude Code memory across sessions explains where each one stops being enough and when a shared store is worth adding.
A one-off task needs a prompt, not an install. If you will not repeat the job next week, a plugin adds a context cost and a set of hooks for nothing.
Hooks you cannot read should stay off. A plugin's hooks run on their own, without you asking each time. If you cannot say what a hook does after reading the file, do not enable the plugin. And if you work only in Codex, autoharness is the wrong choice, because it is a Claude Code plugin. A Codex skill often runs in Claude Code unchanged, and the guide to that covers what carries over and what breaks.
Conclusion
The best open-source Claude Code plugin is the one you would miss on your next task and can read in ten minutes. Start with the official marketplace for your first install, read the Will install list before you accept, and add a community plugin only when a job you repeat needs it. Then check the context cost after a day of real work, not just on the day you install.
The next step is small: pick one job you repeat each week, add one plugin for it, and write down its context cost from the details pane. If the job is about the Claude Code session itself, autoharness is the one of ours that fits. If it is about drafting posts for X or Reddit from your own work, auto-gtm is. For everything else, the Tigerless Labs index lists what we have open.
FAQ
Is an open-source Claude Code plugin safe to install?
Treat it as code you are about to run. Anthropic's documentation says plugins and marketplaces can execute arbitrary code with your user privileges, so read the repository, its license and its hooks, and install only from sources you trust.
Do the Tigerless Labs plugins work in Codex too?
autoharness is a Claude Code plugin. auto-gtm installs in both Claude Code and Codex. A Codex skill often runs in Claude Code without changes, as covered in our guide to carrying a Codex skill across.
Do the Tigerless Labs plugins cost anything?
No. Both are MIT-licensed and install from GitHub. They need no Tigerless account, sign-up or hosted service, and the default auto-gtm setup needs no API key; its Reddit reads use your own logged-in browser session.
Where can I see what a plugin adds before I install it?
Open its details with /plugin install name@marketplace or from the Discover tab of /plugin. The Will install list shows its commands, agents, skills, hooks, and MCP and LSP servers, and the context cost estimate shows the tokens it adds every turn.
How do I measure what my plugins cost in tokens?
cost-xray captures the requests Claude Code and Codex send to the API and attributes tokens and dollars to each source. It observes and reports; it does not cut your spend by itself, and it runs on macOS or Linux, not Windows.
