
elghaly-dev/iris-alexa
Developer toolsЭта страница переведена автоматически. Технические детали смотрите в оригинале. Открыть оригинал
Morning Light Desk MCP — Alexa+ Streamable HTTP MCP (Amazon Developer Hackathon)
Обзор
Self-hosted, read-only Model Context Protocol server for Alexa+ and other MCP hosts. It speaks with the official TypeScript SDK (@modelcontextprotocol/server and @modelcontextprotocol/node) on MCP spec or later. The server stores nothing. It does not log tool arguments. It holds no private key and has no code path that signs, sends, connects, or approves anything. check_link and safety_tip were ported from an earlier private prototype of ours. 1. A pasted 12- or 24-word seed phrase is refused before any rule reads it. 2. The scam rules run on the whole text. A forwarded message that says "approve this transaction to claim your airdrop" or "invest in our staking pool and double your SOL" is a , because those words belong to the sender, not to the person asking Iris. 3. Refusals cover what the person asks Iris to do: price or buy/sell advice, connect or sign in with a wallet, sign or approve. 4.
README
Показан фрагмент оригинального README. Открыть оригинал.
iris-alexa
Amazon Developer Hackathon — Alexa+ track
Self-hosted, read-only Model Context Protocol server for Alexa+ and other MCP hosts. It speaks Streamable HTTP with the official TypeScript SDK (@modelcontextprotocol/server and @modelcontextprotocol/node) on MCP spec 2025-11-25 or later.
git clone https://github.com/elghaly-dev/iris-alexa.git
cd iris-alexa
The server stores nothing. It does not log tool arguments. It holds no private key and has no code path that signs, sends, connects, or approves anything.
✝️🧿🪬
What it does
| Tool | Input | Result |
|---|---|---|
explain_transaction |
Public Solana signature |
Programs, findings, and balance lines from sol-decode.js (copied from our page elghaly-dev/iris-35). One getTransaction call to the public RPC. |
check_scam |
A situation: what the person said, or the message they received |
Fixed rules for nine patterns: seed-phrase request, fake support DM, fake airdrop, doubling giveaway, unlimited approval, authority change (SetAuthority, System Assign), urgent “verify wallet” link, validate/sync-or-be-deactivated, QR code to scan with the wallet. Returns verdict, reason, why, and next_steps. |
check_link |
A url or domain |
Official domain, lookalike domain, punycode host, lure words in the host, or a short link it cannot see through. Never opens the link. |
safety_tip |
Optional topic |
One short reminder written to be spoken aloud. Topics: general, seed_phrase, links, approvals, support, qr_codes. |
clean_up_steps |
target: iphone, android, or wallet |
A checklist you do yourself on that device. |
check_link and safety_tip were ported from an earlier private prototype of ours.
AI-assisted analysis of public pages.
Order of checks
- A pasted 12- or 24-word seed phrase is refused before any rule reads it.
- The scam rules run on the whole text. A forwarded message that says “approve this transaction to claim your airdrop” or “invest in our staking pool and double your SOL” is a scam, because those words belong to the sender, not to the person asking Iris.
- Refusals cover what the person asks Iris to do: price or buy/sell advice, connect or sign in with a wallet, sign or approve.
- Otherwise the verdict is
no_known_patternwith the text “This is not a clearance.”
“Never share your seed phrase” by itself is advice and is not flagged. The same sentence followed by “reply with your 24 word phrase here” is a scam, because the request is still there.
Link rules
- Normalize the host with the WHATWG URL parser (add
https://when needed, IDNA toxn--, lowercase, drop a trailing dot, drop the port). Judgement usesURL.hostnamealone, sohttps://[email protected]/isevil.example.
Установка
npx mcp-inspector --cli http://127.0.0.1:3000/mcp --transport http --method tools/list