EN
Back to the archive

The archive · AI & Models · Strategic decision · 2025–2026

Anthropic open-sources Agent Skills; repo hits 173k stars as rivals adopt it

Anthropic bet giving away its agent-customization standard beats lock-in: OpenAI, Microsoft and a 173k-star repo proved it.

Anthropic

The betThat giving away the SDK for agent customization — skills — wins more than lock-in: whoever defines the standard captures the ecosystem.Live

What the business is

Agent Skills packages specialized know-how (brand guidelines, coding conventions, document rules) into folders an AI loads on demand — Anthropic's answer to making one general agent good at everything.

How it started

In October 2025 Anthropic introduced Skills for Claude: folders of instructions, scripts and resources that load only when relevant — a feature that initially read like a niche developer tool. The engineering post 'Equipping agents for the real world with Agent Skills' appeared 2025-10-16, and the anthropics/skills repo (created 2025-09-22) carried 50+ example skills under Apache-2.0 plus the source-available docx/pdf/pptx/xlsx skills that power Claude's own document features.

What happened

On 2025-12-18 Anthropic announced Agent Skills as an independent open standard with a specification and reference SDK at agentskills.io. Microsoft had already adopted it inside VS Code and GitHub; Cursor, Goose, Amp and OpenCode followed. Ten partners — Atlassian, Figma, Canva, Stripe, Notion, Zapier — launched skills for the directory with no revenue-sharing, and the company said skills ship in Max, Pro, Team and Enterprise plans at no extra cost. OpenAI had quietly replicated the architecture in ChatGPT and its Codex CLI (discovered 2025-12-12 by developer Elias Judin), validating the standard's spread. The announcement landed days after Anthropic donated MCP to the Linux Foundation and co-founded the Agentic AI Foundation with OpenAI and Block.

No ending yet — it is still running.

Background

In October 2025 Anthropic shipped Skills for Claude — folders of Markdown instructions, scripts and resources that an AI loads only when a task needs them. The design principle was 'progressive disclosure': each skill costs a few dozen tokens when summarized in context, with full details streaming in only on demand. That made deep, private customization of a general-purpose agent cheap enough to give away, which is exactly what Anthropic then did.

On December 18, 2025, Anthropic released Agent Skills as an independent open standard with a spec and reference SDK at agentskills.io, plus organization-wide management controls for Team and Enterprise plans. Microsoft had already built support into VS Code and GitHub; Cursor, Goose, Amp and OpenCode adopted it; ten partners including Atlassian, Figma, Canva, Stripe, Notion and Zapier shipped skills at launch with no revenue-sharing. The company's product manager told VentureBeat the GitHub repository had already crossed 20,000 stars.

The strategy's proof came almost immediately from the competition: OpenAI had quietly implemented structurally identical SKILL.md architecture inside ChatGPT and its Codex CLI, discovered by developer Elias Judin on December 12. Unite.AI called the convergence 'the quiet story of the year' — two rival labs agreeing on the same file format, metadata and directory layout for agent capability. The announcement also aligned with Anthropic donating MCP to the Linux Foundation and co-founding the Agentic AI Foundation with OpenAI and Block.

The bet was that ecosystem growth beats proprietary lock-in. By September 2026 the anthropics/skills repo stood at 173k stars and 20.6k forks (GitHub API), the largest open-source artifact of any AI lab, with OpenAI, Microsoft, Google and hundreds of community tools building on the same format.

What has to be true

  • Progressive disclosure made the economics give-away-able: a dozen tokens of context instead of a retraining budget is customization a company can afford to give away.
  • Publishing the spec invited competitors in: once OpenAI and Microsoft adopted SKILL.md, it stopped being Anthropic's feature and became the industry's substrate — a far harder position to displace.
  • Partner skills from Atlassian, Stripe and Figma turned vendor lock-in worries into ecosystem gravity: every integration deepens Claude's usefulness without Anthropic having to build the apps.
  • The same playbook that made MCP a de facto standard (donate, co-found a foundation, invite rivals) told developers the format would outlive any single vendor's roadmap.

What can be applied

When the underlying model gets commoditized, the moat shifts to the customization format: open-sourcing it makes your compatibility the industry's compatibility.

Aftermath

As of 2026-09-03 the anthropics/skills repo is live at 173k stars and a 20k+ fork community, with the Agent Skills spec and reference SDK at agentskills.io, stewarded alongside MCP under the Agentic AI Foundation. Enterprise skill management ships in Team and Enterprise plans; Claude Code was approaching $1B annualized revenue per VentureBeat (Dec 2025). Anthropic's research found 27% of Claude-assisted engineering work would not otherwise have happened, while flagging skill-atrophy and third-party-skill auditing concerns as the ecosystem grows.

Sources

spotted an error? The archive wants to know.

Your turn

You just read one. Describe what you are building, and see who is betting on the same thing.

Free account · 3 free questions · no card

Related cases