Agent Skills
Agent Skills are a lightweight, open format for extending AI agents with specialized knowledge and workflows. A skill is a folder containing a single SKILL.md file plus optional scripts, reference documents, and assets. The SKILL.md file holds a short block of metadata and a set of plain-language instructions that tell an agent how to perform a particular task, from filling PDF forms to following a company's brand guidelines. The format was created by Anthropic, first shipped inside Claude products in October 2025, and then released as a cross-vendor open standard in December 2025 [1][3][11]. On October 22, 2026, Anthropic donated the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, saying that more than 70 agent platforms had integrated it by then, including ChatGPT, Cursor, GitHub Copilot, and Visual Studio Code [6][9].
The basic idea is to keep procedural knowledge outside the model's weights. Instead of retraining a model every time it needs to learn a new workflow, a skill packages that workflow as a version-controlled folder the agent reads on demand [3][4]. This makes capabilities portable: a skill authored once can run on any compatible agent, and a team can keep its conventions in a shared repository or internal marketplace that colleagues use from whichever compatible agent they prefer [3][10].
The SKILL.md format
Every skill lives in its own directory, and the only mandatory file is SKILL.md. The file begins with YAML frontmatter followed by Markdown instructions. The specification published at agentskills.io defines a small set of fields [3]:
| Field | Required | Constraints |
|---|---|---|
name | Yes | 1-64 characters; lowercase letters (a-z), numbers (0-9), and hyphens only; no leading, trailing, or consecutive hyphens; must match the parent directory name |
description | Yes | 1-1024 characters; describes what the skill does and when to use it |
license | No | License name or a reference to a bundled license file |
compatibility | No | 1-500 characters if present; notes environment needs such as the intended product, required packages, or network access |
metadata | No | Map from string keys to string values, for properties the spec does not define |
allowed-tools | No | Space-separated string of pre-approved tools; marked experimental, and support varies between agents |
A minimal skill needs only the two required fields. The Markdown body below the frontmatter has no format restrictions; the spec recommends step-by-step instructions, examples of inputs and outputs, and common edge cases. It also recommends keeping the main SKILL.md under 500 lines, moving detailed reference material into separate files, and keeping file references one level deep from SKILL.md [3].
The field set has not changed since the specification was first published on December 18, 2025. Later edits clarified wording: in 2026 the spec made explicit that name may contain digits, described allowed-tools as a space-separated string, stated that the optional directories are not an exhaustive list, and repeated in its field table that metadata is a map of string keys to string values [22]. As of late October 2026 the specification had no numbered versions or tagged releases [22]. Individual clients can go beyond the shared fields. Anthropic's documentation says Claude Code "extends the standard" with features such as invocation control, running a skill in a subagent, and dynamic context injection [24], and Anthropic's own products add rules of their own, for example barring the reserved words anthropic and claude from skill names [23].
Alongside SKILL.md, a skill can bundle three conventional folders: scripts/ for executable code the agent can run, references/ for documentation it can read when needed, and assets/ for templates, images, or data files. Any other files or directories are also allowed. Files are referenced by relative path from the skill root. A typical layout looks like this [3]:
pdf-processing/
SKILL.md # required: metadata + instructions
scripts/ # optional: executable code
references/ # optional: documentation
assets/ # optional: templates, resources
... # optional: any additional files or directories
The specification's maintainers also publish skills-ref, a reference library whose skills-ref validate command checks that a skill's frontmatter is valid and follows the naming rules [3].
How skills load: progressive disclosure
The mechanism that makes skills practical is progressive disclosure, which loads detail in three stages so that an agent can keep many skills available without flooding its context window [1][3]:
- Discovery. At startup the agent loads only the
nameanddescriptionof every available skill, roughly 100 tokens per skill according to the spec. In Anthropic's implementation this metadata is pre-loaded into the system prompt. It is just enough for the model to know when a skill might be relevant. - Activation. When a task matches a skill's description, the agent reads the full
SKILL.mdbody into context. The spec recommends keeping this under 5,000 tokens. - Execution. The agent follows the instructions, loading bundled files or running scripts only if and when the task calls for them.
Because the heavy material stays on disk until a specific task pulls it in, an agent can keep many skills on hand with only a small context footprint [3]. Anthropic's engineering team adds that agents with a filesystem and code execution tools do not need to read a whole skill into context, so "the amount of context that can be bundled into a skill is effectively unbounded" [1]. In Anthropic's documentation, when Claude runs a bundled script, only the script's output enters the context window, not its code [23].
Relationship to the Model Context Protocol
Agent Skills are frequently compared to the Model Context Protocol (MCP), another open standard that Anthropic introduced in 2024. The two are complementary rather than competing. MCP is a protocol that connects an agent to external tools, data sources, and services, while skills are a packaging format for procedural knowledge about how to carry out a task. When it launched skills, Anthropic said it would "explore how Skills can complement Model Context Protocol (MCP) servers by teaching agents more complex workflows that involve external tools and software" [1]. A skill can describe a multi-step process that calls several MCP tools in sequence, while MCP handles the actual connections. A 2026 survey of the field treats the "complementary roles of skills and MCP" as part of the architectural foundations of skill-equipped agents [4]. The two standards now share a governing body: Anthropic donated MCP to the AAIF in December 2025 and Agent Skills in October 2026 [6][9].
Where skills work
Anthropic shipped Agent Skills across its own products first. At launch on October 16, 2025, skills ran in the Claude apps (for Pro, Max, Team, and Enterprise users), in Claude Code, in the Claude Agent SDK, and on the Claude Developer Platform [1][11]. On December 18, 2025, Anthropic let Team and Enterprise administrators provision skills centrally for their organizations and opened a directory of partner-built skills from Notion, Canva, Figma, Atlassian, and others [7][8]; Zapier, Cloudflare, Sentry, and Vercel also contributed launch statements [8]. Developers on the API use skills through the /v1/skills endpoint, where they run in a sandboxed container without network access or runtime package installation [8][23]. Anthropic's product-specific features are covered in Claude Skills.
Anthropic also maintains a public repository, anthropics/skills, which it describes as "Anthropic's implementation of skills for Claude." It contains a skill template and example skills spanning document creation (docx, pdf, pptx, xlsx), development tasks, and design work. The repository originally held the specification too, but that file now points to agentskills.io. Many of its skills are open source under Apache 2.0, while the document-creation skills are shared as source-available reference implementations rather than open source [2].
Open standard and ecosystem
On December 18, 2025, Anthropic republished the format as an open standard for cross-platform portability, under its own GitHub organization (agentskills) and domain (agentskills.io) [1][6][8]. Anthropic said that within weeks of the October launch, developers had published thousands of skills on GitHub and written loaders for Codex, Cursor, and OpenCode [6].
OpenAI's support predated the open standard. On December 12, 2025, Simon Willison reported that ChatGPT's Code Interpreter had gained a /home/oai/skills folder of document skills and that OpenAI's Codex CLI had merged experimental support for skills about two weeks earlier [14]. Other vendors followed with official support over the next several weeks:
| Product | Developer | Support announced | Details |
|---|---|---|---|
| GitHub Copilot | GitHub | December 18, 2025 | Copilot coding agent, Copilot CLI, and agent mode in VS Code Insiders; picks up skills already set up in .claude/skills [12] |
| OpenAI Codex | OpenAI | December 19, 2025 | CLI and IDE extensions; skills installed per user in ~/.codex/skills or per repository in .codex/skills [15] |
| Visual Studio Code (stable) | Microsoft | January 8, 2026 | Version 1.108, experimental; reads skills from .github/skills or .claude/skills [13] |
| Gemini CLI | January 7, 2026 | Experimental support in preview builds (v0.23.0); enabled by default in releases later that month [17] | |
| Cursor | Cursor | January 22, 2026 | Cursor 2.4, in the editor and CLI [16] |
As of October 30, 2026, the client showcase on agentskills.io listed 46 products. They include JetBrains' Junie, Kiro, Goose, Amp, Factory, Roo Code, OpenHands, Letta, Databricks Genie Code, Snowflake Cortex Code, TRAE, Mistral AI Vibe, Tabnine, Spring AI, Laravel Boost, Pulumi Neo, Nous Research's Hermes Agent, and OpenClaw [3]. Anthropic's own count, given at the October 2026 donation, was more than 70 agent platforms [6]. Because a skill is a folder of files, tools can share them directly: GitHub Copilot and VS Code read the .claude/skills directory that Claude Code uses [12][13].
Registries grew alongside the clients. On January 20, 2026, Vercel released skills, a command-line installer (npx skills add <package>), together with skills.sh, a directory and leaderboard for skill packages. By launch day it had been used to install skills on agents including Amp, Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, Kiro, and OpenCode [18]. By October 2026, Anthropic said 1 million skills had been published and nearly 280 million installed [6]; the AAIF put it as a million skills listed on skills.sh in seven months, installed nearly 280 million times [10].
A February 2026 data-driven study by George Ling, Shanshan Zhong, and Richard Huang analyzed 40,285 publicly listed skills from a major marketplace. It found that skill content is heavily concentrated in software-engineering workflows, with widespread overlap between skills aimed at the same intent and a pronounced supply-demand imbalance across categories [5].
Donation to the Agentic AI Foundation
On October 22, 2026, Anthropic announced that it was donating Agent Skills to the Agentic AI Foundation, the Linux Foundation directed fund it had co-founded with Block and OpenAI in December 2025 and to which it had already given MCP [6][9]. The Linux Foundation announced Agent Skills the same day as one of four new AAIF projects, alongside Open Responses (contributed by OpenAI), OpenEnv (Hugging Face), and OpenSandbox (Alibaba), at the AGNTCon + MCPCon North America conference in San Jose. The additions brought the foundation to 10 hosted projects, up from three at launch [9].
What was donated. Anthropic said it was handing over "the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain" [6]. The GitHub organization holds the specification, its documentation site, and the skills-ref reference library [22]. The AAIF describes the specification as open source under Apache 2.0 and CC BY 4.0, matching the repository's split between code (Apache 2.0) and documentation (CC BY 4.0) [10][22]. Anthropic's separate anthropics/skills repository of example skills is its own implementation and was not among the listed assets [2][6]. Anthropic said developers can keep using the standard or contribute to it without joining the AAIF [6].
Governance. According to Anthropic, "Engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard as individuals, with David Soria Parra and Mahesh Murag as Lead Maintainers" [6]. The AAIF said the project would be guided by "a multi-company Technical Steering Committee" with core maintainers drawn from multiple organizations who serve as individuals [10]. Anthropic proposed an initial roadmap covering "versioned releases, a formal proposal process, and security guidance for people who write and install skills," with feedback on proposed specification changes gathered through the project's Discord [6]. These are proposals: as of October 30, 2026, no versioned release had been published [22].
Statements. David Soria Parra, Anthropic's co-creator and lead maintainer of MCP, said in the Linux Foundation release: "Without the right context, even the most intelligent agents fall short. We launched Agent Skills a year ago to give developers and enterprises alike an open standard to make their agents more capable. Today, we're donating Agent Skills to the AAIF to ensure that anyone building with skills can shape what's next" [9]. Manik Surtani, the AAIF's CTO and co-founder, said: "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10" [9]. Angie Jones, vice president of the AAIF, wrote on the foundation's blog: "That's why a neutral home is the right next step for Agent Skills" [10]. Her post pointed to companies already sharing skills internally: at Block, engineers use an internal marketplace of hundreds of skills with agents including Goose, Codex, Claude, and Amp [10].
Security and governance
Letting an agent read instructions and execute scripts from a downloaded folder carries obvious risks. Anthropic's guidance is to install skills only from trusted sources and to thoroughly audit anything from a less-trusted source, paying particular attention to code dependencies, bundled resources such as scripts, and any instruction or code that tells the agent to connect to potentially untrusted external network sources [1]. Its documentation warns that, depending on what access Claude has, "malicious Skills could lead to data exfiltration, unauthorized system access, or other security risks" [23]. The allowed-tools field is one early attempt to constrain what a skill may do, though the spec marks it experimental [3].
Empirical work has started to quantify the problem. In January 2026, Yi Liu and colleagues analyzed 31,132 skills collected from two marketplaces and reported that 26.1% contained at least one vulnerability, across prompt injection, data exfiltration, privilege escalation, and supply-chain risks. Skills that bundled executable scripts were 2.12 times more likely to be vulnerable than instruction-only skills [19]. A February 2026 survey by Renjun Xu and Yang Yan cited that figure and proposed a Skill Trust and Lifecycle Governance Framework, a four-tier, gate-based permission model that ties a skill's provenance to what it is allowed to do [4]. The data-driven analysis by Ling and colleagues also flagged non-trivial safety risks among skills that can take state-changing or system-level actions [5].
Malicious skills have also appeared in public registries. In early February 2026, Koi Security reported that it had audited all 2,857 skills on ClawHub, the third-party skill registry for OpenClaw, and found 341 malicious ones. It traced 335 of them to a single campaign, dubbed ClawHavoc, that used fake "prerequisite" installs to deliver the Atomic macOS Stealer (AMOS) infostealer [21]. On February 5, 2026, Snyk reported a scan of 3,984 skills from ClawHub and skills.sh: 36.82% (1,467 skills) had at least one security flaw, and human review confirmed 76 malicious payloads [20]. As of October 2026, Anthropic's documentation describes an optional skill content-scanning feature for Claude Enterprise organizations covering custom skills uploaded in claude.ai and Claude Cowork [23]. Security guidance for skill authors and users is one of the three items on the roadmap Anthropic proposed when it donated the standard to the AAIF [6].
Significance
Agent Skills matter because they decouple capability from model training. New behaviors can be added, audited, versioned, and shared as ordinary files, which suits enterprise workflows where teams want consistent, reviewable procedures rather than opaque model updates. The standard also follows the path Anthropic took with MCP: introduced inside its own products, opened to other vendors, and then placed under the Agentic AI Foundation's vendor-neutral governance [6][9]. Neutral governance answers the question of who controls the format. It does not settle the open questions that the AAIF roadmap and the security research point to, such as versioning and the vetting of third-party skills, since the same portability that drove adoption is also what lets an untrusted folder run on dozens of different agents.
Continue reading
- AnthropicAnthropic is an American artificial intelligence (AI) safety and research company founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers, best known for the Claude family of large…Mentioned 35 times above
- Claude (language model)Claude is a family of large language models (LLMs) developed by Anthropic, an American artificial intelligence safety and research company.Mentioned 21 times above
- Cursor (code editor)Cursor is a proprietary AI-assisted code editor and software-development platform made by Anysphere.Mentioned 6 times above
- Agentic AI FoundationThe Agentic AI Foundation (AAIF) is a vendor-neutral, directed fund under the Linux Foundation that provides open governance for the standards and open-source software that AI agents use to connect to tools…Mentioned 4 times above
- Linux FoundationThe Linux Foundation is an American 501(c)(6) nonprofit trade association based in San Francisco, California, that supports the development of Linux and other open-source software projects.Mentioned 4 times above
- GitHub CopilotGitHub Copilot is a hosted artificial intelligence coding assistant developed by GitHub.Mentioned 4 times above
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Anthropic. "Equipping agents for the real world with Agent Skills." Anthropic Engineering, October 16, 2025 (updated December 18, 2025). anthropic.com/...-the-real-world-with-agent-skills
- ^1 ^2Anthropic. "anthropics/skills" repository (Anthropic's example skills; the specification now lives at agentskills.io). GitHub. github.com/...skills
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Agent Skills. "Overview," "Specification," and "Client Showcase." agentskills.io, accessed October 30, 2026. agentskills.io/home , agentskills.io/specification and agentskills.io/clients
- ^1 ^2 ^3Xu, Renjun; Yan, Yang. "Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward." arXiv:2602.12430, February 2026. arxiv.org/...2602.12430
- ^1 ^2Ling, George; Zhong, Shanshan; Huang, Richard. "Agent Skills: A Data-Driven Analysis of Claude Skills for Extending Large Language Model Functionality." arXiv:2602.08004, February 8, 2026. arxiv.org/...2602.08004
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Anthropic. "Donating Agent Skills to the Agentic AI Foundation." October 22, 2026. anthropic.com/...ills-to-the-agentic-ai-foundation
- ^Lardinois, Frederic. "Agent Skills: Anthropic's Next Bid to Define AI Standards." The New Stack, December 18, 2025. thenewstack.io/...-next-bid-to-define-ai-standards
- ^1 ^2 ^3 ^4Anthropic. "Skills for organizations, partners, the ecosystem." Claude blog, December 18, 2025. claude.com/...organization-skills-and-directory
- ^1 ^2 ^3 ^4 ^5 ^6 ^7The Linux Foundation. "Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox." Press release, October 22, 2026. linuxfoundation.org/...ses-openenv-and-opensandbox
- ^1 ^2 ^3 ^4 ^5 ^6Jones, Angie. "Agent Skills: Sharing Team Expertise Across AI Agents." Agentic AI Foundation blog, October 22, 2026. aaif.io/...sharing-team-expertise-across-ai-agents
- ^1 ^2Anthropic. "Introducing Agent Skills." Claude blog, October 16, 2025. claude.com/...skills
- ^1 ^2GitHub. "GitHub Copilot now supports Agent Skills." GitHub Changelog, December 18, 2025. github.blog/...b-copilot-now-supports-agent-skills
- ^1 ^2Microsoft. "December 2025 (version 1.108)." Visual Studio Code release notes, January 8, 2026. code.visualstudio.com/...v1_108
- ^Willison, Simon. "OpenAI are quietly adopting skills, now available in ChatGPT and Codex CLI." Simon Willison's Weblog, December 12, 2025. simonwillison.net/...openai-skills
- ^OpenAI. "Agent skills in Codex." Codex changelog, December 19, 2025. developers.openai.com/...changelog
- ^Cursor. "Subagents, Skills, and Image Generation" (Cursor 2.4). Changelog, January 22, 2026. cursor.com/...2-4
- ^Google. Gemini CLI release notes (v0.23.0, January 7, 2026; v0.25.0, January 20, 2026; v0.26.0, January 27, 2026). geminicli.com/...changelogs
- ^Vercel. "Introducing skills, the open agent skills ecosystem." Vercel changelog, January 20, 2026. vercel.com/...ills-the-open-agent-skills-ecosystem
- ^Liu, Yi; Wang, Weizhe; Feng, Ruitao; et al. "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale." arXiv:2601.10338, January 15, 2026. arxiv.org/...2601.10338
- ^Snyk. "Snyk Finds Prompt Injection in 36%, 1467 Malicious Payloads in a ToxicSkills Study of Agent Skills Supply Chain Compromise." Snyk blog, February 5, 2026. snyk.io/...kills-malicious-ai-agent-skills-clawhub
- ^eSecurity Planet. "Hundreds of Malicious Skills Found in OpenClaw's ClawHub." February 3, 2026. esecurityplanet.com/...-found-in-openclaws-clawhub
- ^1 ^2 ^3 ^4 ^5Agent Skills. "agentskills/agentskills" repository (specification source, commit history, and license). GitHub, accessed October 30, 2026. github.com/...agentskills
- ^1 ^2 ^3 ^4 ^5Anthropic. "Agent Skills" overview. Claude Platform documentation, accessed October 30, 2026. platform.claude.com/...overview
- ^Anthropic. "Extend Claude with skills." Claude Code documentation, accessed October 30, 2026. code.claude.com/...skills
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
7 revisions · v7 · 3,221 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently verified 2026-10-30 against the agentskills.io specification and its commit history, Anthropic and Agentic AI Foundation announcements, vendor changelogs (GitHub, OpenAI Codex, VS Code, Gemini CLI, Cursor) and the cited security research; two minor wording issues corrected.
Cite this page: AI Wiki. "Agent Skills." aiwiki.ai, updated 30 Oct 2026, fact-checked 30 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/agent_skills





