← back

cloudflare/agent-skills-discovery-rfc

GitHub (Cloudflare) · 8 min read · original

Agent Skills Discovery via Well-Known URIs

Status: Draft
Version: 0.2.0
Published Date: 2026-01-17
Updated Date: 2026-03-12

Table of Contents

  1. Abstract
  2. Changelog
  3. Terminology
  4. Problem
  5. Solution
  6. URI Structure
  7. Skill Directory Contents
  8. Progressive Disclosure
  9. Discovery Index
  10. Integrity and Verification
  11. Archive Distribution
  12. Examples
  13. HTTP Considerations
  14. Client Implementation
  15. Security Considerations
  16. Relationship to Existing Specifications
  17. References

Abstract

This document defines a mechanism for discovering Agent Skills using the .well-known URI path prefix as specified in RFC 8615. Skills are currently scattered across GitHub repositories, documentation sites, and other sources. A well-known URI provides a predictable location for agents and tools to discover skills published by an organization or project.

Changelog

v0.2.0

v0.1.0

Terminology

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 RFC 2119 RFC 8174 when, and only when, they appear in all capitals, as shown here.

Problem

Agent Skills give AI agents domain-specific capabilities through structured instructions, scripts, and resources. Today, discovering skills requires:

There is no standard way to answer: "What skills does example.com publish?"

Solution

Register agent-skills as a well-known URI suffix. Organizations can publish skills at:

https://example.com/.well-known/agent-skills/

This provides a single, predictable location where agents and tooling can discover and fetch skills without prior configuration.

URI Structure

Publishers MUST provide an index at:

/.well-known/agent-skills/index.json

Each skill in the index includes a url field pointing to its artifact. While publishers conventionally host skill files under /.well-known/agent-skills/, the url field allows skills to be hosted at any location (e.g., on a CDN or at a versioned path).

Skill names MUST conform to the Agent Skills specification:

Skill Directory Contents

A skill consists of a required SKILL.md file and optional supporting resources:

skill-name/
├── SKILL.md           # Required: instructions + metadata
├── scripts/           # Optional: executable code
│   └── extract.py
├── references/        # Optional: documentation
│   └── REFERENCE.md
└── assets/            # Optional: templates, data files
    └── schema.json

Skills consisting of SKILL.md alone are typically distributed as individual files (type: "skill-md" in the index). Skills with supporting resources are distributed as archives (type: "archive" in the index). See Discovery Index for details.

The SKILL.md file MUST contain YAML frontmatter with name and description fields, followed by Markdown instructions.

Progressive Disclosure

Skills use a progressive loading pattern to manage context efficiently:

Level What When Loaded Token Cost
1 name + description from index At startup or when probing ~100 tokens per skill
2 Full SKILL.md body When skill is activated < 5k tokens recommended
3 Referenced files (scripts, references, assets) On demand, as needed Unlimited

Level 1: Index metadata. Agents fetch index.json to learn what skills exist. Only the name and description are loaded into context initially.

Level 2: Skill instructions. When a task matches a skill's description, the agent fetches the skill artifact. For type: "skill-md" skills, this is SKILL.md directly. For type: "archive" skills, the agent downloads the archive and extracts SKILL.md. The full instructions are loaded into context.

Level 3: Supporting resources. For archive-based skills, the SKILL.md body references additional files via relative links. Agents load these from the unpacked archive on demand as the task requires — a form-filling task might need references/FORMS.md, while a simple extraction task does not.

This pattern means a skill can bundle extensive reference material without paying a context cost upfront. Agents follow links as needed, loading only what the current task requires.

Example: Progressive loading

---
name: pdf-processing
description: Extract text and tables from PDF files. Use when working with PDFs or document extraction.
---

# PDF Processing

## Quick Start

Use pdfplumber to extract text:

```python
import pdfplumber

with pdfplumber.open("document.pdf") as pdf:
    text = pdf.pages[0].extract_text()
```

## Form Filling

For filling PDF forms, see [references/FORMS.md](references/FORMS.md).

## Advanced Table Extraction

For complex tables with merged cells, see [references/TABLES.md](references/TABLES.md) and run `scripts/extract_tables.py`.

An agent handling "extract text from this PDF" loads SKILL.md and stops there. An agent handling "fill out this tax form" follows the link to references/FORMS.md. The table extraction script and reference stay unfetched until needed.

Discovery Index

Publishers MUST provide an index at /.well-known/agent-skills/index.json. The index enumerates all available skills, enabling clients to discover skills in a single request.

Versioning

The index MUST include a top-level $schema field containing a URI that identifies the index schema version. The current schema URI is:

https://schemas.agentskills.io/discovery/0.2.0/schema.json

The $schema URI is an opaque identifier. Clients MUST match it against known schema URIs to determine how to process the index. The URI does not need to be resolvable, though it MAY point to a JSON Schema document that describes the index format.

Clients encountering an unrecognized $schema URI SHOULD warn the user and SHOULD NOT process the index. If $schema is absent, clients SHOULD treat the index as v0.1.0 for backward compatibility. Clients MUST ignore unrecognized fields.

Index Format

{
  "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
  "skills": [
    {
      "name": "code-review",
      "type": "skill-md",
      "description": "Review code for bugs, security issues, and best practices.",
      "url": "/.well-known/agent-skills/code-review/SKILL.md",
      "digest": "sha256:c4d5e6f7..."
    },
    {
      "name": "wrangler",
      "type": "archive",
      "description": "Deploy and manage Cloudflare Workers projects.",
      "url": "/.well-known/agent-skills/wrangler.tar.gz",
      "digest": "sha256:a1b2c3d4..."
    }
  ]
}

The index contains a top-level $schema field and a skills array.

Top-level fields:

Field Required Description
$schema Yes URI identifying the index schema version. See Versioning.
skills Yes Array of skill entries.

Skill entry fields:

Field Required Description
name Yes Skill identifier. MUST conform to the Agent Skills naming specification: 1-64 characters, lowercase alphanumeric and hyphens only, no leading/trailing/consecutive hyphens.
type Yes Distribution type. MUST be "skill-md" (single SKILL.md file) or "archive" (bundled archive).
description Yes Brief description of what the skill does and when to use it. Max 1024 characters per the Agent Skills spec. SHOULD match the description field in the skill's SKILL.md frontmatter.
url Yes URL to the skill artifact. For type: "skill-md", this points to the SKILL.md file. For type: "archive", this points to the archive file. See URL Resolution.
digest Yes SHA-256 content digest of the artifact at url, formatted as sha256:{hex} where {hex} is 64 lowercase hexadecimal characters. See Integrity and Verification.

[!NOTE]
In a future version, url may become optional for type: "skill-md" entries, defaulting to /.well-known/agent-skills/{name}/SKILL.md.

URL Resolution

The url field specifies where to fetch the skill artifact. URLs are resolved per RFC 3986 Section 5 using the index URL as the base URI. URLs may be:

For type: "skill-md", url conventionally follows the pattern /.well-known/agent-skills/{name}/SKILL.md, though publishers MAY use any URL.

For type: "archive", url points to the archive file. Clients SHOULD determine the archive format from the server's Content-Type header, falling back to the URL file extension if the header is absent or generic (e.g., application/octet-stream). See Archive Distribution.

Clients encountering an unrecognized type value SHOULD skip that skill entry and MAY warn the user.

Backward Compatibility

The v0.2.0 index format is not backward-compatible with v0.1.0. Key differences:

Clients MUST check the $schema field to determine how to process the index. An unrecognized $schema URI indicates the index structure may have changed incompatibly; see Versioning.

Integrity and Verification

All digests in the index use SHA-256 and are formatted as sha256:{hex}, where {hex} is 64 lowercase hexadecimal characters.

The digest field on each skill entry is the SHA-256 hash of the raw bytes of the skill's artifact:

sha256:{SHA-256(raw_bytes)}

Verification

Clients MUST verify downloaded content against the digest in the index. A mismatch indicates the content is corrupted or has been tampered with; clients MUST NOT use unverified content.

Archive Distribution

Skills with supporting files (scripts, references, assets) are distributed as archives with type: "archive" in the index. The archive contains the full skill directory, including SKILL.md and all supporting resources.

Supported Formats

Archives SHOULD be in .tar.gz (gzip-compressed tar) or .zip format. Clients MUST support at least .tar.gz and .zip. Each format has different tradeoffs:

Archive Structure

The archive contents represent the skill directory — files are placed at the archive root, not nested inside a wrapper directory.

Archive Safety

After verifying the archive's digest (see Integrity and Verification), clients unpacking an archive MUST:

  1. Reject archives containing path traversal sequences (..) or absolute paths.
  2. Reject archives containing symlinks or hard links that resolve outside the skill directory.
  3. Consider enforcing a reasonable limit on total unpacked size to prevent denial-of-service via decompression bombs.

Distribution Guidance

Simple skills — those with only SKILL.md — SHOULD use type: "skill-md". Archives are intended for skills with supporting files where a single download is more efficient and preserves directory structure, file permissions, and symlinks.

Examples

Simple skill (SKILL.md only)

A minimal skill contains just the required SKILL.md:

GET /.well-known/agent-skills/git-workflow/SKILL.md

---
name: git-workflow
description: Follow team Git conventions for branching and commits.
---

# Git Workflow

Create feature branches from `main`:

```bash
git checkout -b feature/my-feature main
```

Commit messages use conventional commits format:

```
feat: add user authentication
fix: resolve null pointer in login
docs: update API reference
```

Complex skill with resources

A skill with scripts and reference documentation, distributed as an archive:

wrangler.tar.gz (archive contents)
├── SKILL.md
├── scripts/
│   ├── deploy.sh
│   └── init.sh
├── references/
│   ├── COMMANDS.md
│   └── CONFIGURATION.md
└── assets/
    └── wrangler.toml.template

The SKILL.md references these files for progressive disclosure:

---
name: wrangler
description: Deploy and manage Cloudflare Workers projects.
---

# Wrangler

## Deployment

Run `scripts/deploy.sh` to deploy your Worker.

For available commands, see [references/COMMANDS.md](references/COMMANDS.md).
For configuration options, see [references/CONFIGURATION.md](references/CONFIGURATION.md).

Discovery index

{
  "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
  "skills": [
    {
      "name": "code-review",
      "type": "skill-md",
      "description": "Review code for bugs, security issues, and best practices.",
      "url": "/.well-known/agent-skills/code-review/SKILL.md",
      "digest": "sha256:c4d5e6f7..."
    },
    {
      "name": "git-workflow",
      "type": "skill-md",
      "description": "Follow team Git conventions for branching and commits.",
      "url": "/.well-known/agent-skills/git-workflow/SKILL.md",
      "digest": "sha256:a7b8c9d0..."
    },
    {
      "name": "wrangler",
      "type": "archive",
      "description": "Deploy and manage Cloudflare Workers projects.",
      "url": "/.well-known/agent-skills/wrangler.tar.gz",
      "digest": "sha256:f1e2d3c4..."
    }
  ]
}

HTTP Considerations

Servers MUST:

Servers SHOULD:

Clients MUST:

Client Implementation

Clients discovering skills from a well-known endpoint MUST:

  1. Fetch index.json. Retrieve /.well-known/agent-skills/index.json to enumerate available skills.

  2. Check schema version. Match the $schema field against known schema URIs. If absent, treat the index as v0.1.0. Clients SHOULD NOT process an index with an unrecognized $schema URI and SHOULD warn the user. Clients MUST ignore unrecognized fields.

  3. Use digests for caching. Compare each skill's digest against locally cached values. If the digest matches, the skill is unchanged and the client MAY skip re-downloading.

  4. Fetch and verify skill artifacts. For skills that need updating:

  5. For type: "skill-md": Download SKILL.md from the skill's url. Compute its SHA-256 and verify against digest.
  6. For type: "archive": Download the archive from the skill's url. Compute its SHA-256 and verify against digest. Unpack the archive and validate its structure (see Archive Safety).

  7. For an unrecognized type: Skip the skill entry and warn the user.

  8. Apply progressive disclosure. Load only name and description at discovery time. Load SKILL.md when a skill is activated. For archive-based skills, load supporting resources (scripts, references, assets) on demand as the task requires.

  9. Cache aggressively. Skills change infrequently. Respect Cache-Control headers and consider caching content for the duration of a session. Use skill digests to invalidate cached content when updates are detected.

  10. Gate script execution. Clients SHALL NOT execute files under scripts/ by default. Clients SHALL consider implementing a permissions model that only executes scripts bundled with a skill when explicitly allowed by the user or client configuration. Consider sandboxing execution environments and restricting filesystem and network access. Never execute scripts from untrusted origins without user approval.

Security Considerations

The security considerations from RFC 8615 Section 4 apply. Additional considerations for skills:

Relationship to Existing Specifications

This document builds on:

References

ai.news — daily digest