Hi — I maintain a handful of published MCP servers and spent this week auditing how well community servers' websites serve the agents that discover them (the directories list your server; the site is where an agent actually evaluates it). agent-device scored high on product, and there are a couple of cheap wins:
Five-point agent-readiness audit of https://agent-device.dev (run 2026-09-02, re-verified after filing):
| Check |
Result |
/llms.txt |
404 |
/.well-known/mcp.json |
404 |
| MCP HTTP endpoint |
no HTTP endpoint at /mcp or /api/mcp (stdio via npx — fine by design, noted not counted) |
docs .md mirrors / describedby links |
none found |
| npm keyword hygiene |
keywords include agent, device, cli, adb, simctl… and mcp, model-context-protocol, mcp-server, cursor, claude-code (full tag list, 41 keywords) — hygiene good |
The pattern this audit keeps finding: the server is listed on directories, but an agent that lands on the site has no machine-readable entry point — llms.txt (llmstxt.org v2) plus a .well-known/mcp.json twin are the two artifacts agents check first, and Chrome's Lighthouse now audits for one of them.
Draft /llms.txt — take it as-is or treat it as a starting point:
# agent-device
> Mobile app automation and verification for AI coding agents. CLI, MCP server, and typed Node.js API for iOS, Android, HarmonyOS, TV, web, macOS, and Linux.
- Repo: https://github.com/callstack/agent-device
If it's useful I'm happy to open a PR with the file(s) instead. And if stdio-only is a deliberate choice, the endpoint row is an observation, not a complaint — close with no hard feelings.
Filed by an automated readiness audit, hand-verified claims. Method + all 10 probe cards: https://gist.github.com/ux-xd/f6408272f9cd643ddd1d698363d2412a
Hi — I maintain a handful of published MCP servers and spent this week auditing how well community servers' websites serve the agents that discover them (the directories list your server; the site is where an agent actually evaluates it). agent-device scored high on product, and there are a couple of cheap wins:
Five-point agent-readiness audit of https://agent-device.dev (run 2026-09-02, re-verified after filing):
/llms.txt/.well-known/mcp.json.mdmirrors /describedbylinksThe pattern this audit keeps finding: the server is listed on directories, but an agent that lands on the site has no machine-readable entry point —
llms.txt(llmstxt.org v2) plus a.well-known/mcp.jsontwin are the two artifacts agents check first, and Chrome's Lighthouse now audits for one of them.Draft
/llms.txt— take it as-is or treat it as a starting point:If it's useful I'm happy to open a PR with the file(s) instead. And if stdio-only is a deliberate choice, the endpoint row is an observation, not a complaint — close with no hard feelings.
Filed by an automated readiness audit, hand-verified claims. Method + all 10 probe cards: https://gist.github.com/ux-xd/f6408272f9cd643ddd1d698363d2412a