docs: cross-tool entry points (AGENTS.md, Cursor rule, mcp.json example)

AGENTS.md

@@ -0,0 +1,21 @@
+# AGENTS.md — arcade-eval
+
+Cross-agent entry pointer (Cursor, Claude Code, and other agent tools honor this file).
+This repo is the tool-agnostic source of truth for the Arcade.dev MCP-gateway evaluation.
+
+## Before doing anything
+1. `git pull` (on rejection, `git pull --rebase`). `cp config/.env.example .env` and fill creds.
+2. Read, in order: **`README.md` → `STATUS.md` → `LIVE-POC.md` → `GROUND-RULES.md`**.
+3. Run the **live-state check** (in `GROUND-RULES.md`) before trusting the live instance.
+
+## Working rules (full text in GROUND-RULES.md)
+- Write only inside **your** `categories/catN-*/` subtree + your own `STATUS.md` section.
+  Shared files (`config/targets.yaml`, `lib/`, top-level docs) are append-mostly; `git pull
+  --rebase` before push.
+- **Credentials live only in `.env`** — never commit them.
+- **Never write the criteria Google Doc from a session** — compose `criteria-section-N.md`
+  locally; the human pastes. Criterion wording is verbatim.
+
+## Connecting an MCP client
+Gateway URL pattern: `https://api.arcade.st.dev/mcp/{slug}`. Cursor users: see
+`.cursor/mcp.json.example`. Claude Code users: `claude mcp add --transport http …`.