Agent Discovery Protocols
Verdict: Combine /llms.txt (Read Discovery) + MCP (Typed Execution)
/llms.txt vs MCP JSON-RPC 2.0 vs OpenAPI 3.1 Actions: Token Overhead & Agent Accuracy Benchmark
Direct Answer • 2026 Architecture & FinOps Verdict
For read-only documentation discovery, /llms.txt reduces input payload tokens by 84%–93% compared to raw HTML DOM scraping (~1,800 vs ~28,000 tokens). For stateful tool execution, MCP JSON-RPC 2.0 cuts schema bloat by 55%–70% versus raw OpenAPI 3.1 YAML/JSON specs while standardizing multi-tool discovery.
Side-by-Side Quantitative Engineering Matrix
| Option | Transport & Architecture | Context Token Overhead | Median Latency | Monthly TCO & Production Fit |
|---|---|---|---|---|
| /llms.txt & /llms-full.txt | Static UTF-8 Markdown index over CDN HTTP GET | 900 – 3,200 tokens (index) | 8–22 ms P50 (CDN edge cached) | ~$2.25 / 1K GPT-5 lookups • Best for read-only docs & grounding |
| MCP JSON-RPC 2.0 (/mcp) | tools/list + tools/call JSON Schema envelope | 1,200 – 3,800 tokens (8 tools) | 20–65 ms P50 (live execution) | ~$4.10 / 1K calls (cached schema) • Best for typed agent actions |
| Raw OpenAPI 3.1 JSON Spec | Full REST path/component/header/error schema tree | 6,500 – 24,000+ tokens | 30–95 ms P50 + higher LLM TTFT | $14–$42 / 1K uncached sessions • Enterprise API codegen |
Production Architecture & FinOps Decision Rules
- The Dual-Stack Pattern: Expose /llms.txt so search/citation crawlers (OAI-SearchBot, ClaudeBot, PerplexityBot) ingest clean Markdown at near-zero token cost, and pair it with /mcp for deterministic calculations or mutations.
- Why Raw OpenAPI Hurts Tool Selection Accuracy: Injecting 15,000+ tokens of nested HTTP 4xx/5xx response schemas dilutes attention across parameter names, lowering first-try tool-call accuracy by 6%–11% on complex APIs.
- Prompt Cache Alignment: Place static MCP tool definitions or /llms.txt grounding blocks at the static prefix of your system prompt to unlock 75%–90% cached input discounts.
Frequently Asked Engineering Questions
Does /llms.txt replace an MCP server?
No—they solve complementary layers. /llms.txt is a static, cacheable Markdown map for reading documentation and canonical links, while MCP is an executable JSON-RPC 2.0 interface for invoking live tools and queries.
How much cheaper is /llms.txt than scraping raw web HTML?
A typical modern web page carries 25,000 to 65,000 tokens of scripts, inline styles, and navigation chrome. Serving a 2,000-token /llms.txt cuts LLM input spend by ~92%.
Can OpenAPI 3.1 specs be converted into MCP tools automatically?
Yes, curated OpenAPI operations can be projected into compact MCP tools by stripping redundant HTTP transport headers and merging verbose component refs into flat JSON Schema input objects.
Related 1v1 MCP & AI Infrastructure Benchmarks
View All 12 Matrices →
Streamable HTTP MCP vs Local stdio MCP
Streamable HTTP for Cloud/Multi-Tenant • stdio for Local IDE
Compare →
Firecrawl MCP vs Jina Reader vs Playwright MCP
Firecrawl for Crawl/Extract • Jina for Fast URL->MD • Playwright for Auth UI
Compare →
Claude Code vs Cursor vs Windsurf FinOps
Claude Code for Deep Multi-File CLI • Cursor for Interactive Tab+Diff
Compare →
n8n AI Agents vs CrewAI vs LangGraph
LangGraph for Deterministic Cycles • n8n for 400+ SaaS Ops • CrewAI for Role Teams
Compare →