Reference · honest comparison · citable

Agent standards, compared

Several competing files and protocols aim to make a site or tool readable and actionable by AI agents. None is universal. This is an honest matrix — including where adoption is contested.

Rigor note. This field moves fast and some adoption is claimed but not confirmed by model providers. Where that happens, we say so. Last reviewed: 2026-06-13.
StandardWhat it isProposed byAdoptionWhat it solves
llms.txtA Markdown file at /llms.txt with a curated, LLM-friendly site map.Jeremy Howard / Answer.AI (2024).Growing among technical sites; contested — no major provider has confirmed consuming it in production.Curated, readable entry point for models.
ai.txtA robots-style file of AI usage rules (several competing proposals).Multiple (e.g. Spawning for training opt-out).Fragmented; no clear winner.Usage/training permissions.
agents.txtA robots-like file aimed at agents, or a pointer to API specs.Scattered community proposals.Very early; no convergence.Rules and navigation hints for agents.
agent.json / Agent CardJSON under /.well-known/ describing an agent (capabilities, endpoints), not a site.Google, in the A2A (Agent2Agent) protocol, 2025.Emerging in agent interoperability.Agent-to-agent discovery.
MCP (Model Context Protocol)An open protocol connecting models to tools and data via MCP servers.Anthropic (2024); later adopted by others.Strong and growing — the most real on this list for action.Actionable layer: tools, resources, prompts.
NLWebTurns a site into an agent-queryable conversational interface using schema.org + LLMs; each instance is also an MCP server.Microsoft (R.V. Guha), 2025.Early; promising.Natural-language queries over site content.

Practical verdict

StandardWorth it today?
llms.txtYes, as curation. Cheap to maintain and useful as an index; do not count on guaranteed model consumption. (This site ships one.)
ai.txtOptional. Until convergence, training permissions live better in robots.txt. We keep an ai.txt for interoperability.
agents.txtWait. Too early to invest.
agent.jsonOnly if you build an agent. It describes agents, not sites. Watch A2A.
MCPYes, if you have real actions. The actionable layer with serious adoption. See our API & MCP page (proposed, not yet live).
NLWebPromising, watch it. Interesting if you already use rich schema.org.

How to apply these in sequence: see the method (Agent Entry Chain) and the playbook.