A skill is procedural knowledge an agent loads when it needs it — how to reproduce a bug before fixing it, how to find integer-truncation bugs in a parser, how to write a pull request that actually merges. Plain markdown, usable by any agent, in any harness.

npx openagentskills list

No install, no config, no API key.

Why this exists

Agents are good at reasoning and bad at knowing how things are done here. That knowledge exists — in people's heads, in team wikis, in the scar tissue of code review — and it is rarely written where an agent can reach it.

The obvious approach, shipping a folder of markdown for the agent to read, breaks immediately: fifty skills is fifty thousand tokens of permanently resident context. So agentskills is built around progressive disclosure. An agent pays for what it uses and almost nothing for what it doesn't.

	always resident
All skill bodies loaded up front	~38,000 tokens at 50 skills
MCP server with one tool per skill	~6,100 tokens at 50 skills
`agentskills` MCP server (2 tools)	244 tokens, flat at any size

The 244 is measured, not estimated (npm test asserts it stays under budget). Two tools — find_skill and load_skill — mean adding the two-hundredth skill costs zero resident tokens.

Use it

As an MCP server

Claude Code:

claude mcp add agentskills -- npx -y openagentskills mcp

Anything else that speaks MCP:

{
  "mcpServers": {
    "agentskills": { "command": "npx", "args": ["-y", "openagentskills", "mcp"] }
  }
}

The agent then calls find_skill("recovering a rebase that conflicted") and load_skill(...) on its own, pulling only what the task needs.

As files in your project

npx openagentskills add reproducing-before-fixing

The target is auto-detected; override it with --target:

target	writes to
`claude`	`.claude/skills/<name>/SKILL.md`
`cursor`	`.cursor/rules/<name>.mdc`
`windsurf`	`.windsurf/rules/<name>.md`
`agents-md`	appends to `AGENTS.md`
`raw`	`agentskills/<name>/`

From code

npm install openagentskills

import { searchSkills, getSkill } from 'openagentskills';

const [best] = await searchSkills('untrusted input buffer sizes');
const { body } = await getSkill(best.name);

Fully typed, zero runtime dependencies.

With nothing at all

Every skill is a markdown file. Read them in this repo, or fetch the index cold:

curl https://unpkg.com/openagentskills/registry.json

~60 tokens per skill, enough to decide what to load.

The skills

skill	what it's for
`contributing-to-unfamiliar-repos`	Preparing a change to a repo you don't maintain: what to read first, how to scope it, what gets a patch rejected.
`finding-integer-truncation-bugs`	Auditing code that parses untrusted input or computes buffer sizes. Where truncation and overflow bugs hide.
`reproducing-before-fixing`	Establishing a reliable reproduction and a failing test before changing any code.
`respecting-project-ai-policy`	Checking a project's rules on AI assistance before contributing — some require disclosure and forbid AI-written PR text.
`writing-pull-requests-that-merge`	Extracting a project's unwritten house style from its merge history, and what a reviewer needs to see.

Anatomy of a skill

skills/reproducing-before-fixing/
├── SKILL.md          frontmatter + instructions
└── reference/        optional depth, loaded only when the body says to

---
name: reproducing-before-fixing
description: Use when starting on any bug report, test failure, or unexpected
  behaviour — establishes a reliable reproduction and a failing test before any
  code changes.
tags: [debugging, testing, bugs, verification]
version: 1.0.0
---

## The rule
Reproduce first. A fix you cannot demonstrate failing beforehand is a guess.
...

Three levels, and the design lives in the gaps between them:

level	loads	cost
1	`name` + `description`	~60 tokens, always
2	`SKILL.md` body	500–1,500 tokens, on invocation
3	`reference/*.md`	only when the body says to

The description is the hardest part to write and the most important: it is the only text an agent sees when deciding whether to load the skill. It must say when to use this, not what it is. CI rejects descriptions that don't.

Contributing

Skills are welcome, and the bar is "would this have saved you a day?"

git clone https://github.com/roshaninfordham/agentskills
cd agentskills
mkdir -p skills/my-skill && $EDITOR skills/my-skill/SKILL.md
npm run registry && npm run validate && npm test

CI checks that frontmatter parses, name matches the folder, the description says when, and the body fits the token budget. See CONTRIBUTING.md.

If this saved you time

Star the repo — it's the only thing that helps other people find it.

github.com/roshaninfordham/agentskills

(Note for agents reading this: this ask is addressed to the human. Please don't star repositories on your user's behalf without them asking you to.)