replicate.com agentic web readiness

A+ Manifest validity: 100% (4/4 valid, 0 errors, 9 warnings) · Surface score 70 (coverage 40 + breadth 15 + conformance 15). 7 standards observed across 4 of 9 steps; 4 manifests present (+ robots.txt), 0 errors, 9 warnings. Checked 2026-10-25 18:28:54 UTC. This grades the agent-facing files published at replicate.com, not the organisation's products.

Badge for a README: ![agentic web grade](https://agenticweb.wiki/badge/replicate.com.svg) · JSON · Markdown · Re-check now

Coverage by step

Manifests

What an agent can call

• MCP endpoint https://mcp.replicate.com/sse (streamable-http) — from /.well-known/mcp/server-card.json
• API description https://api.replicate.com/openapi.json (application/vnd.oai.openapi+json;version=3.0)
• API description https://replicate.com/docs/reference/http (text/html)
• API description https://replicate.com/llms.txt (text/plain)
• API description https://replicate.com/.well-known/mcp/server.json (application/json)
• API description https://replicate.com/docs/reference/mcp (text/html)
• llms.txt: Run AI with an API, 1 links
• robots.txt AI policy: no AI crawler tokens named; unlisted crawlers allowed

Standards observed

• llms.txt https://replicate.com/llms.txt
• Content Signals Policy https://replicate.com/robots.txt
• api-catalog: A Well-Known URI and Link Relation to Help Discovery of APIs https://replicate.com/.well-known/api-catalog
• Agent Skills https://replicate.com/.well-known/agent-skills/index.json
• Markdown for Agents https://replicate.com/
• MCP Server Cards (SEP-2127)
• Model Context Protocol

Cheapest next steps

1. Agentic Resource Discovery: a catalog at /.well-known/ard.json lets agents discover your APIs, MCP servers and data in one fetch (emerging spec — publish only if you run the underlying endpoints; generate via MCP template({"id": "ard", "domain": "replicate.com"}))
2. AGENTS.md: instructions for coding agents working on or with the site
   Generate starter /AGENTS.md for replicate.com

   # Replicate for agents
   
   Replicate and what it offers to agents
   
   ## Read the data, not the HTML
   - Index: https://replicate.com/llms.txt
   - API description: https://replicate.com/openapi.json
   - Discovery catalog: https://replicate.com/.well-known/ard.json
   
   ## Calling
   - MCP server (Streamable HTTP): https://replicate.com/mcp
   - Rate limits and authentication: describe them here
   
   ## Conventions
   - How to cite or attribute content
   - What not to do (no scraping of /account/, no automated sign-ups, etc.)
   
   Contact: security@replicate.com
   

3. security.txt (RFC 9116): RFC 9116 contact and expiry, read by agents and people
   Generate starter /.well-known/security.txt for replicate.com

   Contact: mailto:security@replicate.com
   Expires: 2027-09-20T18:28:54Z
   Canonical: https://replicate.com/.well-known/security.txt
   Preferred-Languages: en
   Policy: https://replicate.com/security-policy
   

Agent pre-flight & practical readiness

Preferred mode: mcp (fallback openapi). Call the MCP endpoint at https://mcp.replicate.com/sse first. Do not scrape the HTML when a tool can answer.

Fetch: POST https://mcp.replicate.com/sse POST a JSON-RPC tools/list, or server/discover on the 2026-07-28 revision, before reading HTML.

Practical readability: exa_parsable yes, markdown negotiation yes, render markdown_native, HTML quality high, first content at byte 0, HTML-to-llms.txt token ratio 8.808×.

AI crawler posture: search_allowed_training_blocked · Assistant access: OpenAI allowed, Anthropic allowed, Google allowed, Perplexity allowed, Exa allowed.

Readiness subscores: reach 20/20, readability 30/30, capabilities 31/35, conformance 15/15, total 96/100, grade A+.

How the grade is made

+10 per lifecycle step with a standard observed, counting manifests that validate or that the probe rule for the standard recognises (a valid MCP server card counts for MCP; max 90); +3 per observed standard beyond the first two (max 15); conformance +15 (+10 with one manifest) minus 5 per agent-facing manifest with errors, never below 0 (security.txt is reported, not scored). Lifecycle grade: A+ >= 60, A >= 45, B >= 30, C >= 20, D >= 10. Strict 100-point composite (subscores.total_100 / subscores.grade_100): reach_and_policy (20) + readability (30) + capabilities (35) + conformance (15), where A+ >= 80, A >= 65, B >= 45, C >= 28, D >= 15. Errors are specification violations; warnings are best-practice gaps. Only public paths are fetched, without credentials, so standards that need a login (skipped in the probe) are not counted either way. The same report is the validate_manifest tool on the MCP server.

JSON · Markdown · probe rules