Trust Vault MCP Server
Exposes 8 read-only Trust Vault tools over MCP (Streamable HTTP) for querying protocol overview, tokens, market rates, orders, platform stats, fees, and currencies. Enables on-chain reads and static config without wallet or signing.
README
Trust Vault MCP Server — Phase 1 (Read-Only)
Exposes 8 read-only Trust Vault tools over MCP (Streamable HTTP transport):
get_protocol_overview, get_supported_tokens, get_market_rates,
list_open_orders, get_order_status, get_platform_stats,
get_fee_structure, get_currencies_and_processors.
No wallet, no signing, no auth surface — pure on-chain reads + static config.
Setup
npm install
Set env vars (or a .env if you wire one in):
TRUST_EXPRESS_IDL_PATH=/path/to/your/trust_express.json # your real Anchor IDL
SOLANA_RPC_URL=https://api.devnet.solana.com # or mainnet
TRUST_VAULT_KNOWLEDGE_PATH=/path/to/knowledge/trustvault.md # same file /api/chat already loads
PORT=3939
npm run build
npm start
Before this goes live — 3 things to fix, all flagged inline in code
-
src/tools/fetchOrders.ts—pricePerTokenscale is unconfirmed. The program docs say "fiat per whole token, raw fiat units" but don't state whether NGN is stored as whole naira or kobo (×100). Currently passed through raw. Check againstBuyOrderCard/SellOrderCarddisplay logic in the client and fix that one line — it's load-bearing for every tool that surfaces a price. -
src/program.ts— pointTRUST_EXPRESS_IDL_PATHat your real IDL. Deliberately not faked. The server refuses to start (with a clear error) if the file is missing, and separately refuses to start if the IDL'saddressfield doesn't match the known program ID inconstants.ts— catches a devnet/mainnet mix-up before it causes silent wrong reads. -
src/constants.ts—FEE_STRUCTURE. Set to the corrected 50% LP / 30% platform / 20% validators split. Thetrust-vaultskill file still says 40/40/20 — that's stale, don't copy from it. Fix the skill file separately so future docs don't drift again.
Known issue — needs a look before you rely on multi-request sessions
While testing the Streamable HTTP session handling (src/server.ts), I hit
a case where tools/list returned "Server not initialized" on the request
immediately following a successful initialize — i.e. the onsessioninitialized
callback's transports.set(id, transport) didn't appear to be visible yet
on the next request in the same process. I didn't get to root-cause this
before switching to just handing off the files — it may be a real race in
how server.connect() vs transport.handleRequest() are sequenced here,
or it may have been sandbox process weirdness (multiple stale node
processes potentially fighting over the port during testing). Worth
re-testing this cleanly in your own environment — if it reproduces there
too, the app.post("/mcp", ...) handler in src/server.ts is where to dig
in, specifically whether onsessioninitialized fires and populates the
transports map before the next HTTP request's lookup runs.
Design notes
- Every tool is read-only — no
.methods()calls, only.account.fetch()/.all(). The Anchor provider has no real signer (see comment insrc/program.ts) — it throws if anything ever tries to sign, as a deliberate guardrail. get_order_statusdeliberately does NOT fetchValidatorVotePDAs (votes_for/votes_against/executed) — that's a separate account keyed offkeccak256(payout_reference), out of scope for general-info tools. Seesrc/tools/orderStatus.tsfor the reasoning.get_protocol_overviewreads the sameknowledge/trustvault.mdyour/api/chatroute already loads — update one file, both surfaces pick it up.
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。