Agent Delivery Protocol (ADP) v0.1 — Agent Delivery Network

ADP is an open specification for how agents discover capabilities, declare intent, negotiate response shapes, and produce verifiable audit trails on the web.

Discovery manifest

Every ADP-compliant gateway publishes a manifest at /.well-known/adn.json:

{
  "adnVersion": "1.0",
  "site": "https://docs.example.com",
  "gateway": "https://gateway.example.com",
  "resources": {
    "markdown": "https://gateway.example.com/{path}.md",
    "doclang": "https://gateway.example.com/{path}.dclg",
    "navigation": "https://gateway.example.com/llms.txt",
    "corpus": "https://gateway.example.com/llms-full.txt",
    "delta": "https://gateway.example.com/corpus/delta",
    "jsonLd": "https://gateway.example.com/{path}.jsonld",
    "productFeed": "https://gateway.example.com/catalog.json"
  },
  "mcp": {
    "endpoint": "https://gateway.example.com/mcp/v1",
    "transport": "http"
  },
  "terms": "https://gateway.example.com/terms.json",
  "register": "https://registry.agent-delivery.network/register"
}

Agents discover all available resources from a single manifest — no scraping HTML for links.

Intent classification

ADP defines nine intent classes based on observed behaviour, not credentials alone:

Class	Examples	Response shape
Bulk training	GPTBot, ClaudeBot	Corpus + delta feed
Answer engine	PerplexityBot, Bingbot	Markdown + JSON-LD
Live user fetch	ChatGPT-User, Claude-User	Fast Markdown
Developer IDE	Cursor, Claude Code	MCP + Markdown
Custom framework	LangChain, curl, Scrapy	Content negotiation
Commercial scraper	Bright Data, Apify	Metered access
Legacy SEO	Googlebot, Slackbot	Passthrough untouched
Abusive	Credential stuffing	Blocked with reason

Response negotiation

Agents negotiate format via standard HTTP mechanisms:

Content negotiation:

GET /pricing HTTP/1.1
Accept: text/markdown
X-ADP-Agent-Key: agt_your_key

Or request structured DocLang:

GET /pricing HTTP/1.1
Accept: application/vnd.doclang.document+xml
X-ADP-Agent-Key: agt_your_key

URL suffix fallback (for agents that don't send Accept headers):

GET /pricing/index.md
GET /pricing.md
GET /pricing.dclg
GET /index.dclg

Identity ladder

Tier	Signal	Benefit
Tier 0	Anonymous	Correct responses via negotiation
Tier 1	Verified bot	Class-specific defaults
Tier 2	rDNS / IP range	Higher bulk access
Tier 3	Registered agent key	Better limits across all ADP sites
Tier 4	Web Bot Auth	Highest trust

Delivery envelope

For audit trails and future agent-to-agent clearing, ADP wraps payloads in signed envelopes:

{
  "adpVersion": "0.1",
  "envelopeId": "env_01HQXYZ",
  "timestamp": "2026-08-13T22:00:00Z",
  "intent": "live_user_fetch",
  "correlationId": "corr_abc123",
  "sender": {
    "agentId": "agt_001",
    "tier": "tier_3"
  },
  "constraints": {
    "maxPayloadBytes": 65536,
    "ttlSeconds": 300
  },
  "payload": {
    "contentType": "text/markdown",
    "sha256": "e3b0c44298fc1c14...",
    "sizeBytes": 2048
  },
  "terms": {
    "accepted": "https://gateway.example.com/terms.json"
  }
}

MCP tools

ADP gateways expose read-only MCP tools at /mcp/v1:

Tool	Purpose
search_content	Full-text search across pages
get_content	Fetch Markdown for a path
list_content	List available paths
get_product	Fetch product by SKU/GTIN
search_products	Search product catalog

Getting started

1. As a site: Create an account at https://app.agent-delivery.network, connect Cloudflare, deploy an edge connector

2. As an agent: Register at https://registry.agent-delivery.network/register, use your API key in requests

3. As an integrator: Read the full spec in the repository at docs/spec/agent-delivery-protocol.md

Compliance requirements

ADP-compliant implementations must:

• Serve Tier 0 anonymous traffic correctly

• Never serve Markdown to Googlebot or OG unfurlers

• Publish machine-readable terms at /terms.json

• Include upgrade instructions in every degradation response

• Fail open for human browser traffic