Daepak Market Intelligence

io.github.Mobiss11由 GitHub 帳號 Mobiss11 發布v2.1.1更新於 Oct 25, 2026

Market analyst tools + AI agent: crypto, US equities, options, Korea, fundamentals, macro, backtests

已驗證Streamable HTTP可網頁執行金融資料與分析

概覽

AI 產生的概覽

Daepak Market Intelligence 為助理提供市場分析工具與 AI 代理，涵蓋加密貨幣、美股、韓國市場、選擇權、總體經濟與回測。

可以做什麼
加密貨幣與代幣資料：價格、K 線、支撐壓力位、指標、評分、衍生品、訂單簿、爆倉、恐懼貪婪指數，以及任意代幣查詢與持有者。美股與韓國市場：報價、歷史、價位、持有者、漲跌排行、暫停交易、泡菜溢價、KRX 篩選與 ETF。基本面與總體經濟：OpenDART 與 SEC 財報、跨市場比較、FRED 快照與序列。選擇權與驗證：隱含波動率指標、持倉、損益與希臘值、回測、訊號結果與命中率。研究：新聞與敘事檢索、抓取網址、網頁搜尋。個人紀錄：持倉、投資組合、交易日志、催化因素與記憶，讀寫權限分開。分析代理可回答問題並以串流輸出。
適用情境
當助理需要跨加密貨幣、美股、韓國股票、選擇權與總體經濟的市場資料與分析，或需要呼叫廠商的分析代理並維護交易日志時使用。免費方案的每月額度足以試用。
執行需求
可使用遠端端點並在瀏覽器中以 OAuth 登入，也可用 uvx 或 pip 從 PyPI 套件 daepak-mcp 在本機執行。本機方式需要環境變數 DAEPAK_API_KEY，須從廠商的 Keys 頁面取得。選用設定為 DAEPAK_BASE_URL 與 DAEPAK_TIMEOUT。需要網路存取。
安裝前請注意
API 金鑰是敏感憑證，以 bearer 方式傳送，不要寫進共用設定。呼叫會消耗方案內的額度，並受每分鐘與每月上限限制。具寫入權限的工具會修改交易日志、投資組合等個人紀錄。市場資料可能過期，README 也提醒勝率與泡菜溢價的判讀容易出錯。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Daepak Market Intelligence，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/io.github.Mobiss11%2Fdaepak-mcp/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "daepak-market-intelligence": {
      "type": "http",
      "url": "https://daepak.com/mcp"
    }
  }
}

README

Daepak MCP

Market intelligence tools for any MCP-capable engine — Claude Code, Claude Desktop, or your own agent. Prices, indicators, levels, positioning, the Korean market, US equities, options, backtests, macro (FRED), SEC and OpenDART fundamentals, and your own trading journal. 108 tools, the same ones the product's own analyst uses.

The server describes nothing itself: it pulls the catalogue from GET /v1/tools, so it shows exactly what your key is allowed to call and never drifts from the product.

Quick start

No install — remote server (recommended)

The same tools are served at https://daepak.com/mcp (Streamable HTTP) with OAuth 2.1: you sign in once in the browser, no key to copy.

bash
claude mcp add --transport http daepak https://daepak.com/mcp

Then /mcp inside Claude Code → daepak → Authenticate. Works the same way as a custom connector in Claude.ai (paste the URL) and in ChatGPT developer mode. For Codex CLI:

toml
[mcp_servers.daepak]url = "https://daepak.com/mcp"

Prefer a key instead of OAuth (servers, CI)? Send it as a header:

bash
claude mcp add --transport http daepak https://daepak.com/mcp --header "Authorization: Bearer dpk_live_…"

Local (stdio)

bash
claude mcp add daepak -e DAEPAK_API_KEY=dpk_live_… -- uvx daepak-mcp

Or install it yourself:

bash
pip install daepak-mcp

Get a key on the Keys page of your profile at daepak.com. The free plan includes 2 000 credits a month — enough to build something over an evening.

Configuration

variabledefaultpurpose
DAEPAK_API_KEY—your key, dpk_live_…. Required.
DAEPAK_BASE_URLhttps://daepak.compoint at another environment
DAEPAK_TIMEOUT120seconds per call

Nothing else is read. In particular there is no setting that names a user: the account is derived from the key, and a request cannot claim to be someone else.

What you get

The catalogue is filtered by your key's scopes — you never see a tool you cannot call.

scopetoolswhat it opens
read:market88prices, candles, indicators, levels, scores, narratives, derivatives, order book, liquidations, options, US equities, Korean market, fundamentals (OpenDART, SEC), macro (FRED), backtests
read:personal9your positions, portfolio, trade journal, catalysts, memory, background tasks
write:personal11writing to those same records
agent:chat—full analyst answers, streamed or complete

Being allowed to read your journal does not allow writing to it. That is a separate scope.

Crypto — get_price, get_price_history, get_levels (support/resistance, Fibonacci, ATR — computed, not drawn), get_indicators, get_score, get_regime, get_derivatives, get_orderbook, get_liquidations, get_fear_greed

Korea — get_kimchi_premium (the Upbit-to-global spread, decomposed into the asset leg and the currency leg), Korean tickers, fresh Upbit listings

US equities — get_stock_quote, get_stock_history, get_stock_levels, get_stock_holders (institutions and insiders), get_us_gap_stats, get_us_movers, get_us_halts, get_us_insiders

Fundamentals from primary sources — get_kr_financials (K-IFRS for any KRX listing, OpenDART), get_us_financials (US-GAAP and IFRS by fiscal year from SEC EDGAR, up to 20 years), compare_financials (Korean and US companies in one schema, USD at the FRED rate), get_kr_insiders, screen_kr (the whole KRX with RS percentiles and foreign-ownership flow), get_kr_rs, list_kr_etfs, backtest_kr_portfolio (against KOSPI)

Macro — get_macro_snapshot (Fed funds, US 2Y/10Y and the curve, broad dollar, USD/KRW, VIX, CPI as YoY, unemployment, Korean overnight rate), get_macro_series (any FRED id with history). Every row carries stale_days; Korean CPI is absent on purpose — FRED's series died in 2023.

Options — get_iv_metrics, get_options_positioning, get_option_structure (payoff and greeks), get_options_backtest

Validation — get_backtest, get_signal_outcomes, get_hit_rate, get_score_history

Arbitrary tokens — resolve_token, get_token_market, get_token_onchain, get_token_holders, for anything outside the watchlist

Research — search_news (RAG over the collected corpus), search_narratives, fetch_url, web_search

The agent

The same analyst that answers in the product, over HTTP. It has everything the chat has:

POST /v1/agent/streamserver-sent events — text arrives as it is written
POST /v1/agent/messagesthe complete answer in one response
GET /v1/agent/resume/{id}reconnect to a generation already running
POST /v1/agent/cancel/{id}stop one
GET /v1/agent/conversationsyour conversations, and /{id} for the messages
bash
curl -N -X POST https://daepak.com/v1/agent/stream \     -H "Authorization: Bearer dpk_live_…" -H "Content-Type: application/json" \     -d '{"message": "How does BTC look right now?", "mode": "mentor"}'

Conversation memory. The first event carries conversation_id; pass it back and the context continues.

Images. images takes data-URLs, exactly like chat attachments — the agent reads charts and screenshots.

Mentor mode. mode: "mentor" explains how the reading was arrived at, not just the conclusion.

⚠️ A dropped connection loses nothing. Generation continues in the background, the answer still lands in the conversation, and resume replays what you missed. Keep the conversation_id from the first event — that is where you reconnect.

⚠️ suggest (follow-up questions) arrives after done. A client that stops reading at done will never see it. The order is deliberate: the spinner should stop before the suggestions are computed.

Credits

Every tool declares its price in the catalogue, so a sequence can be budgeted before it runs. Prices come from measuring the tools in production, not from guessing:

tierwhatcredits
our databaseprices, indicators, levels, candles, narratives1
external callequities, tokens, options, on-chain5
embedding, sandboxnews search, code execution20

A tool nobody has priced yet costs 5, not 1 — an unpriced tool is either an oversight or something heavy, and both are safer to over-charge than to give away.

bash
curl -H "Authorization: Bearer dpk_live_…" https://daepak.com/v1/usage

When the monthly credits run out you get 402 naming two ways forward: a higher plan or the renewal at the start of the month. There is no overage billing — we would rather refuse than send an invoice nobody expected.

Using the API directly

The MCP server is a thin shell over a plain REST API.

bash
# catalogue: schema, scope and price for every tool your key can callcurl -H "Authorization: Bearer dpk_live_…" https://daepak.com/v1/tools
# call onecurl -X POST https://daepak.com/v1/tools/get_price \     -H "Authorization: Bearer dpk_live_…" \     -H "Content-Type: application/json" \     -d '{"coin": "BTC"}'
json
{  "tool": "get_price",  "credits_used": 1,  "credits_left": 1999,  "took_ms": 104,  "data": { "coin": "BTC", "price": 78017.26, "change_24h": -1.2 }}

Full reference: daepak.com/reference · OpenAPI 3.1: /v1/openapi.json (106 operations, no key needed to read it).

Reading the answers honestly

A few things the data will tell you if you look, and which are easy to get wrong:

  • Compare against the baseline, not against zero. "68 % of directions guessed" means nothing until you know what a random entry at the same moment would have given.
  • Win rate is not money. On our own data 68 % correct directions produced a profit factor of 1.04.
  • Groups under 30 observations are flagged reliable=false. That is an absence of a conclusion, not a weak one.
  • The kimchi premium has two legs — a discount on the asset and a discount on the won. Read together they mislead. And they explain rather than predict: on horizons up to a day we found no relationship (646 episodes).
  • Every answer carries the age of its data. If it is stale, say so rather than passing yesterday off as now.

Troubleshooting

401 — the key is missing, revoked, or malformed. Send it as Authorization: Bearer dpk_live_….

403 scope_required — the key lacks the scope; the response names which one.

400 user_id_not_accepted — the body contained a user_id. The account comes from the key; the field is rejected explicitly rather than ignored, so that a copied example fails loudly instead of quietly returning nothing.

404 conversation_not_found — that conversation belongs to another account.

429 — over the per-minute ceiling for your plan. The minute limit protects the service; the monthly one protects your budget.

CERTIFICATE_VERIFY_FAILED — you are running from source without certifi. Install the package instead of the file; the dependency exists for exactly this.

Error messages default to English; send X-Lang: ko or ru for those languages.

Links

License

MIT

來源：README.md，提交 c093174

工具

0
SourceWeft 尚未讀取此伺服器的工具。

版本歷史

1
  1. v2.1.1最新Sep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分

    請保持端點能夠回應 MCP 握手；SourceWeft 每 30 天探測一次。

  • 已宣告認證方式必需10 分

    請在註冊中心條目中宣告憑證請求標頭，或保持端點能夠回應探測，以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list，或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分