The gap between human-readable documentation and machine-readable discovery is widening. A new scan of four major developer platforms conducted on November 22, 2026, reveals that while the industry giants have mastered the basics of serving content to AI crawlers, they are failing to expose the critical endpoints needed for autonomous agents to actually use their APIs. The report, published on DEV.to by independent developer manhourplan, highlights a stark contrast: Vercel, Supabase, and Cloudflare score well on static content delivery but miss the mark on Model Context Protocol (MCP) and x402 payment discovery files.
The Scoreboard: Basics vs. Machine Discovery
The scanner evaluated sites based on ten specific signals, ranging from server-rendered HTML to /.well-known/mcp.json. Vercel.com scored 74/100, missing llms-full.txt, mcp.json, and x402.json. Supabase.com performed slightly better with a 79/100, but still lacked the essential mcp.json and x402.json files. Cloudflare took the top spot among the major platforms with an 84/100, though it too missed the x402 payment catalog and the RFC 9727 API catalog. The only site to achieve a perfect 100 was the authorβs own platform, manhourplan.com, which serves as both the test subject and the tool provider.
Why llms.txt Is Not Enough
All three major platformsβVercel, Supabase, and Cloudflareβpassed the 'basics' test. They serve real HTML without requiring JavaScript execution, include llms.txt files, and explicitly define robots.txt rules for AI crawlers like GPTBot, ClaudeBot, and PerplexityBot. However, the author argues that these measures only ensure an AI can read the documentation, not that it can interact with the platform. The missing /.well-known/mcp.json file is the standard way for agents to discover an MCP server, allowing them to call tools programmatically without human intervention. Similarly, the absence of /.well-known/x402.json means agents cannot automatically discover pay-per-call APIs using USDC.
Quick Wins for Dev Platform Builders
For developers running their own platforms, the report suggests immediate, low-effort fixes. Adding a simple llms.txt file at the root directory, containing a one-line description and links to key pages, is described as the highest-value action, taking only about ten minutes. Additionally, server-rendering content is non-negotiable; if a homepage loads as an empty
Key Takeaways
- Vercel scored 74/100, Supabase 79/100, and Cloudflare 84/100 on AI agent readiness.
- None of the three major platforms exposed /.well-known/x402.json for pay-per-call API discovery.
- None of the three major platforms (Vercel, Supabase, Cloudflare) exposed /.well-known/mcp.json for Model Context Protocol discovery; only the author's own site, manhourplan.com, achieved this.
- llms.txt and server-rendered HTML are now table stakes, but machine-readable API catalogs remain rare among major platforms.
The Bottom Line
If your platform wants to be called by agents, not just read by them, you need to stop treating llms.txt as the finish line and start publishing your MCP and x402 manifests. The infrastructure for autonomous commerce is already here, but most of the industry is still hiding their keys.