agent-ready — https://aadhar.sh origin https://aadhar.sh ── doors a machine can walk through ──────────────────────────────────────── open llms.txt 16598 bytes shut markdown twin text/html open agent card 7039 bytes open api catalog 4127 bytes shut MCP server answered no content-type, not JSON 3 of 5 open ── what this cost to build (counted 2026-08-05) ──────────────────────────── capability files lines buys ──────────────────── ───── ────── ────────────────────────────────────────── MCP server 3 574 tools/list + tools/call, stateless 2026-0… llms.txt 1 208 a readable map, hand-written agent discovery 1 113 agent-card, api-catalog, anonymous auth markdown twins 2 918 every prose page, generated so it cannot … surface registry 3 792 one source for nav, sitemap, agent catalog bot identity 2 319 RFC 9421 signing on every outbound fetch x402 paywall 1 140 machine payment for the full corpus frame renderer 1 307 the 80-column TUI representation the tools 5 1932 finger, lens, dict, cache, radar total 19 5303 Baseline agent compatibility — an MCP server, an llms.txt, and the discovery files — is 895 lines of the 5303. That part is a weekend. The rest is drift-proofing and the tools themselves, which is where the cost actually hides and where nobody budgets. The harder half is not in this table at all: it is the discipline of not lying to a caller. Every genuinely difficult problem here was an honesty problem, not a protocol one — a door that could not be checked reported as shut, a score computed without its inputs, a delta that was really plain zstd, an ETag that can never match. A site can be perfectly conformant and still mislead every agent that visits. doors 3/5 counted, never scored — a check that could not run is not a failure