Olaverse Lab

io.github.Olaverse-Labs由 GitHub 帳號 Olaverse-Labs 發布v0.2.1更新於 Nov 7, 2026

Document extraction with calibrated confidence, decisions, language ID in 608 languages, diacritics.

已驗證Streamable HTTP可網頁執行知識與記憶資料與分析AI 與機器學習

概覽

AI 產生的概覽

Olaverse Lab 將小型專用語言模型包裝成工具：依範本把文件抽取成帶信賴度的 JSON、型別化判斷、語言辨識與變音符號還原。

可以做什麼
  • 依範本把文字、JPEG/PNG 圖片或 PDF 抽取成結構化 JSON，並為每個欄位提供校準後的信賴度；長文件以背景工作執行，稍後取回結果
  • 針對訊息、工單或紀錄回答型別化問題（是/否、選項、評分），並給出校準機率
  • 在 608 種語言中辨識文字語言，可選流量或覆蓋模式
  • 為 11 種語言還原聲調符號、重音符號與阿拉伯語母音符號
  • 產生某段文字所能回答的搜尋式問題，並使用該文字本身的語言
適用情境
適合需要從收據、表單、掃描件或 PDF 擷取欄位，並依逐欄位信賴度判斷可信度的情境，也適合處理英文與非洲語言文字的語言辨識、變音符號還原或問題產生。
執行需求
需要一組在 Olaverse 主控台建立的 API 金鑰：本機 uvx 套件透過 OLAVERSE_API_KEY 傳入，託管端點則透過 Authorization 的 Bearer 標頭傳入。本機執行需要 uv/uvx，託管端點免安裝。可選設定 OLAVERSE_API_URL 與 OLAVERSE_JOB_WAIT。沒有文字層的掃描 PDF 需以 JPEG 或 PNG 圖片送出。
安裝前請注意
每次呼叫都依你的 API 金鑰計費，超過贈送額度後會產生費用。API 金鑰屬於機密，透過 OLAVERSE_API_KEY 或 Authorization 標頭傳遞。送出的文件、訊息與文字會傳到 Olaverse 的伺服器。信賴度僅對模型所針對的文件類型經過校準，自動化前請用自己的樣本驗證門檻。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Olaverse Lab，將其新增到工作區。
  2. 連接 Authorization token 憑證。它們會被私密儲存，絕不會在公開頁面上展示。
  3. 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.Olaverse-Labs%2Folaverse-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": {
    "olaverse-lab": {
      "type": "http",
      "url": "https://mcp.olaverselab.com/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

README

Olaverse MCP server

Olaverse Lab's API as tools for AI agents: Claude, Cursor and any other MCP client. The models are small, task-specific and strongest on English and African-language text.

ToolWhat it does
extract_documentA document (text, a JPEG/PNG photo or scan, or a PDF) to JSON shaped like your template, with a calibrated confidence for every field. Long documents are split, extracted in parallel and merged.
get_extraction_jobCollects a long-document result that's still running.
decideTyped questions about a message, ticket or record (yes/no, choice, score), with calibrated probabilities. 19 languages.
detect_languageThe language of a text, among 608 (built African-first, good on very short text).
restore_diacriticsPuts back tone marks, accents and Arabic vowel marks in 11 languages, including Yorùbá, Igbo, Hausa and Arabic.
generate_questionsSearch-style questions a passage answers, in the passage's own language (25 languages).

Set up

You need an Olaverse API key: create one at console.olaverselab.com (new accounts get $1 of free credit). Calls are billed to your key at the usual prices.

The server runs with uv (uvx downloads and runs it).

Claude Code

bash
claude mcp add olaverse --env OLAVERSE_API_KEY=your-key -- uvx olaverse-mcp

Claude Desktop, Cursor and others (the mcpServers block of the client's MCP config file)

json
{  "mcpServers": {    "olaverse": {      "command": "uvx",      "args": ["olaverse-mcp"],      "env": { "OLAVERSE_API_KEY": "your-key" }    }  }}

Hosted (no install). Olaverse also runs this server at https://mcp.olaverselab.com/mcp (streamable HTTP). Send your key with each request as Authorization: Bearer your-key. The hosted server takes documents as text, or as an image or pdf in base64; it doesn't read files or fetch URLs.

Use it

Ask in plain words, and the agent picks the tool:

  • "Extract the merchant, date, total and line items from ~/Downloads/receipt.jpg."
  • "Is this customer message asking for a refund, and how urgent is it? «Ẹ jọ̀ọ́, wọ́n ti yọ owó lẹ́ẹ̀mejì…»"
  • "Which language is 'Mo fẹ́ lọ sí ọjà'?"
  • "Add the tone marks back: se eranko naa si gbo o?"

Confidence. extract_document returns a confidence for every field. It's calibrated: fields at 0.95 and above are right about 95% of the time on the documents it was built for. Check a threshold on a sample of your own documents before you automate on it.

Settings

VariableDefault
OLAVERSE_API_KEYrequiredYour API key.
OLAVERSE_API_URLhttps://api.olaverselab.com
OLAVERSE_JOB_WAIT900Seconds a long-document extraction waits before handing back a job id for get_extraction_job.

Scanned PDFs (no text layer) aren't read directly: send their pages as JPEG or PNG images.

Licence

Apache-2.0.

mcp-name: io.github.Olaverse-Labs/olaverse-mcp

來源：README.md，提交 b5d4b2a

工具

6

extract_document

Turn a document into JSON shaped like `template`, with a calibrated confidence (0-1) for every field. template: the shape of the answer; each leaf names a type: "verbatim-string" (copied exactly), "string", "number", "integer", "date" (YYYY-MM-DD), "date-time", "boolean", ["a", "b"] (pick one), [["a", "b"]] (pick any), ["string"] (a list), or [{...}] (a list of objects, e.g. line items). Example: {"merchant": "verbatim-string", "total": "number", "date": "date", "items": [{"name": "verbatim-string", "price": "number"}]} Give the document as exactly one of: text: the document as text. image: a JPEG or PNG, base64 or a data URL (a photo, scan or screenshot of one page). pdf: a PDF with a text layer, base64. file_path: a JPEG/PNG, PDF or text file on this machine (local server only). instructions: optional guidance, e.g. "dates are day/month/year" or "amounts in kobo". Long documents (over 60,000 characters) are split into parts, extracted in parallel and merged; that can take minutes. Missing values come back as null. Treat fields with confidence below about 0.95 as needing a check.

輸入參數pdftextimagetemplate必填file_pathinstructions

get_extraction_job

The result of a long-document extraction job that extract_document left running (by its id, exj_...).

輸入參數job_id必填

decide

Answer typed questions about a message, ticket, record or form, with a calibrated probability for each answer. state: the thing to decide about: a string (e.g. a customer message) or a JSON object (a record). questions: {question_id: {...}}, each one of: {"type": "noul", "instructions": "The customer is asking for a refund."} -> probability it's true {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "payments", "delivery": "orders"}} {"type": "score", "instructions": "How urgent?", "criteria": ["can wait", "today", "immediately"]} Works in 19 languages, including English, Pidgin, Hausa, Yorùbá, Igbo, Swahili and Arabic.

輸入參數state必填questions必填

detect_language

Identify the language of a text among 608 languages (built African-first; good on very short text). mode: "traffic" (user messages and chat; favours languages people actually use) or "coverage" (all languages equal, for building corpora). Returns the ISO 639-3 code, script, name, confidence and runners-up.

輸入參數modetext必填top_k

restore_diacritics

Put back diacritics, tone marks or Arabic vowel marks in text typed without them. language: yor (Yorùbá), ibo (Igbo), hau (Hausa), vie, pol, tur, por, spa, fra, ita or ara (Arabic). Example: "se eranko naa si gbo o?" (yor) -> "ṣé ẹranko náà sì gbọ́ ọ?"

輸入參數text必填language必填

generate_questions

Write n natural, search-style questions that a passage answers, in the passage's own language (25 languages). Useful for FAQs, study cards and search or RAG training data. passage: up to 8,000 characters.

輸入參數npassage必填language

版本歷史

1
  1. v0.2.1最新Nov 7, 2026

評分

已滿足適用檢查項的 100%

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
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分