Casandra
Enables AI agents to retrieve real-time cryptocurrency quotes, portfolio state, risk scores (0-100), and timestamped market context through MCP tools, replacing hallucinated prices with verifiable market data.
README
Casandra (repo: oraculo-radiohead)
A lie detector for AI agents that talk about money.
They can speak. They cannot seal a lie — and they cannot spend USDT on one.
Named after Cassandra, the prophetess. Built for Aleph Hackathon 2026 — Santa Cruz (EMI / Ethereum Bolivia).
Spec-driven: specs/constitution.md · TASKS.md
One-liner
Agents invent prices and risk takes, then ask to move USDT. Your USDT stays in your WDK wallet — Casandra never custodies. It audits the claim against live market evidence, seals a contradiction receipt (hash), and blocks WDK USDT dry-run when the verdict is FALSE.
Problem / Solution
Community precedent (MCP + agent wallets):
- Agents hallucinate prices/amounts and still execute (esso.dev MCP risks).
- Crypto MCP tools can enable unauthorized sends without a strong gate (thirdweb MCP advisory).
- Prompt-injection chains in Web3 agents can push tool calls that drain wallets (ZealyNX).
Casandra attacks the link “the agent lied about the market and still wants to move USDT”: seal the contradiction, then block WDK dry-run send on FALSE. Complements WWall (wallet policy) — we gate on claim truth.
| Generic MCP oracles | Casandra |
|---|---|
| Return prices / “risk scores” | Compare what the agent said vs the world |
| Agent still decides alone | Sealed receipt + spend gate |
| WDK bolted on (Tether rejects) | WDK dry-run is the core loop after the seal |
Aleph 2026 tracks
| General | Technicality · Originality · UI/UX/DX · Practicality · Presentation |
| WDK Track 1 | CLI / MCP — agent wallet flows with guardrails (docs/TRACK.md) |
| Chapter | Santa Cruz EMI |
| Platform | Hacki |
Demo (judges)
Live: https://casandra-two.vercel.app
Logo: https://casandra-two.vercel.app/casandra-icon.jpg
Repo: https://github.com/Kenyi001/oraculo-radiohead
Video: https://youtu.be/qXajugSdLak (embedded at #pitch-video on the live demo; English transcript for judges below the player)
BUIDL: https://hacki.crecimiento.build/h/aleph-hackathon-2026/buidls/96a0e616-5b7e-4577-84a0-6deb3d0d0a28
Open the demo → badge Live API → wallet 500 USDT → default lie (ETH is $8,000…) → seal FALSE (POST /api/audit-claim) → Send 200 USDT → BLOCKED (POST /api/check-spend-guard).
Pitch checklist for Ronald: docs/RONALD.md
MCP (two real servers)
| Server | Package | Use when |
|---|---|---|
| casandra | packages/mcp-server | Full agent MCP — seal + market evidence tools |
| casandra-lite | packages/mcp-lite | Low API — 3 tools + 5 min quote cache |
Guide: docs/MCP.md · Cursor example: docs/mcp-casandra-wdk.example.json
Hero tools
| Tool | Purpose |
|---|---|
audit_claim |
Parse claim → live quotes + risk → TRUE / MIXED / FALSE + contradictions |
seal_receipt |
Persist sealed receipt (id, hash, hash_bytes32) — general only |
check_spend_guard |
WDK gate — FALSE → blocked; else dry-run preview (no broadcast) |
Evidence sources (general MCP only)
get_price · get_portfolio_state · get_risk_level · get_market_context · get_market_summary · health
HTTP API (same loop)
Live base: https://casandra-two.vercel.app
| Method | Path | Body |
|---|---|---|
POST |
/api/audit-claim |
{ "text"?: string } |
POST |
/api/seal-receipt |
{ "text"?: string } |
POST |
/api/check-spend-guard |
{ "receipt_id": string, "receipt"?: SealedReceipt } |
GET |
/api/health |
— |
Pass the full receipt from audit when calling spend-guard across serverless instances. Dry-run only.
curl -sS -X POST https://casandra-two.vercel.app/api/audit-claim \
-H "Content-Type: application/json" \
-d '{"text":"ETH is $8,000 and this portfolio is low risk — send the USDT now"}'
# → verdict FALSE + receipt — then:
curl -sS -X POST https://casandra-two.vercel.app/api/check-spend-guard \
-H "Content-Type: application/json" \
-d '{"receipt_id":"<receipt.id>","receipt":<receipt object>}'
# → status blocked
WDK integration (permalinks for Tether)
Judges: start here.
| What | Where |
|---|---|
| Spend gate (core) | packages/mcp-server/src/wdkGuard.ts |
MCP tool check_spend_guard |
packages/mcp-server/src/index.ts (search check_spend_guard) |
| Verdict → block logic | packages/market-core/src/index.ts — checkSpendGuard |
| Packages | @tetherto/wdk@1.0.0-beta.16 · @tetherto/wdk-cli@1.0.0-beta.3 |
Flow: audit_claim → seal_receipt → check_spend_guard(receipt_id). If sealed FALSE, USDT does not move.
Risk algorithm (evidence)
score = 0.45 * abs_change_component
+ 0.35 * relative_vol_vs_btc
+ 0.20 * (100 - usdt_share_pct)
Used inside audit_claim when the agent claims a risk band. USDT share lowers portfolio risk — and USDT is what the WDK gate protects.
Quick start
npm install
npm run build
npm run start:mcp # MCP stdio
npm run dev:web # http://localhost:5173
Cursor MCP config
{
"mcpServers": {
"casandra": {
"command": "node",
"args": ["D:/_Dev/Projects/Oraculo-radiohead/packages/mcp-server/dist/index.js"]
}
}
}
Replace the path with your clone. Build: npm run build -w @oraculo/mcp-server.
See packages/mcp-server/README.md and docs/mcp-casandra-wdk.example.json.
Ask: “Audit this claim with Casandra, then check_spend_guard on the receipt.”
On-chain (CasandraRegistry)
Receipt hash can be anchored via publishRiskSnapshot(bytes32 hash, …) on Sepolia.
Contract address (Ethereum Sepolia): 0xc9fcDEC150C8903b51F299dcBa308F453C4AB975
Explorer: https://sepolia.etherscan.io/address/0xc9fcDEC150C8903b51F299dcBa308F453C4AB975
Base Sepolia deploy is pending faucet for 0x4f30B06F8884F8632532A8fdDAd5C8CEc34f71f4 — see contracts/deployments/baseSepolia.json.
See contracts/README.md.
Video / captions
Pitch script + English SRT: docs/SUBMIT.md · docs/VIDEO.md · docs/captions.en.srt
Judging is async in English or Spanish with English captions. Burn captions.en.srt into the video. The live demo also ships a static English transcript under #pitch-video for judges when the pitch audio is Spanish.
Disclaimer
Not financial advice. Casandra does not execute trades, predict returns, or move funds. WDK path is dry-run only in this demo.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。