Stuga

dev.stugav0.1.10更新於 Oct 2, 2026

Agent edits await review by default in a self-hosted workspace for documents and databases.

已驗證Streamable HTTP可網頁執行DatabasesKnowledge & MemoryProductivity & Workflow

概覽

AI 產生的概覽

Stuga 是自架的檔案與資料庫工作區，AI 助理的修改會以修訂形式出現，預設要由人審核並接受後才生效。

功能
代理透過 MCP 連線到自架節點，節點中存放協作檔案與以 SQLite 為基礎的資料庫。預設情況下，它們的寫入會以修訂形式等待人工接受或拒絕，檔案或資料庫的擁有者或工作區管理員也可以選擇讓 AI 修改直接生效。每次執行都會記錄代理、它所代表的人、用戶端與模型以及每一項變更，代理還能查詢哪些段落是自己撰寫的。檔案支援即時協作編輯、留言與版本歷史，資料庫則提供具型別的資料表、已儲存的檢視與唯讀 SQL。
適用情境
當團隊希望 AI 助理在共用檔案與資料表中工作，但在變更生效前保留人工把關時，適合使用。適合重視代理變更的審核、歸屬與稽核，而非即時生效的自架情境。
執行需求
需要一個透過串流 HTTP 存取的 Stuga 自架節點，位址為你的節點位址。可安裝在搭載 Apple 晶片、macOS 13 或以上版本的 Mac，也可在 x86-64 或 arm64 上以 Docker 安裝；節點由一個 Node.js 程序與包含 pgvector 和 pg_search 的 Postgres 組成。建立帳號與工作區後，透過瀏覽器登入或 API 金鑰連接代理。節點本身不需要 AI 金鑰，因為每個代理自帶模型。
安裝前請注意
代理可以變更檔案與資料庫，因此審核設定決定變更是立即生效還是等待審核。登入授權會把應用程式限定在所选工作區，並限定為唯讀或可讀取並建議變更；API 金鑰可限定到資料夾、設為唯讀或設定到期日，不再需要時應撤銷存取。重新命名、移動、刪除、分享與審核設定仍由人掌握。每次 MCP 呼叫（包括讀取）都會記入稽核記錄，節點僅在使用相關功能（例如 AI 供應商）時發出對外請求。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Stuga，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。

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "stuga": {
      "type": "http",
      "url": "https://{stuga_host}/mcp"
    }
  }
}

README

[圖片]

Stuga

AI suggests. You decide.

Self-hosted documents and databases where edits from Claude Code, Codex or any MCP client arrive as tracked changes and, by default, land only when a person accepts them. Use it on your own or with your team.

Download for Mac (Apple silicon, macOS 13+) · Install with Docker

Install, then create your account and workspace and connect an agent. Help for people using Stuga: stuga.dev/docs.

[Claude Code's edits to a document arriving as tracked changes, accepted and rejected one by one]

How agent edits work

  • Review by default. An agent's changes wait for a person, who is notified and accepts or rejects each one in the document or table. Nothing lands on a timer.
  • Nobody is blocked. Collaborators never see pending agent text and keep editing. The agent carries on, and its later reads include its own pending edits.
  • Or apply at once. The owner of a document or database, or a workspace admin, can choose Let AI edits apply directly. Changes then land at once, recorded, attributed and revertible.
  • A run per session. Each run records the agent, the person it acts for, the client and model it reported, and every change. Review AI edits lists the runs that need someone.
  • Who wrote this? An agent can ask a document which passages agents wrote and whether a person accepted them.
  • Instructions that stack. The workspace, folders, documents and databases can each carry instructions for agents. They advise the model; permissions and review decide what a write does.

[Claude Code's changes to a table waiting as proposals, then accepted]

What's in it

  • Documents written together. Real-time editing (Tiptap over a Yjs CRDT) with presence, comments, @mentions and version history.
  • Databases with SQL. Typed tables beside your documents, each database its own SQLite file, with saved views, row pages and CSV import. Agents, Ask and the REST API run read-only SQL on them.
  • Search that follows permissions. Keyword (BM25, pg_search) and semantic (pgvector) search in one SQL statement, with access checked inside it. Chinese and Japanese are split into words.
  • Built-in AI, off until set up. A co-author and a table assistant, whose edits are reviewed like an agent's, and Ask, which cites sources. OpenAI, Anthropic, Gemini, DeepSeek and more, or Ollama.
  • Sharing and sign-in. Workspaces, folders, per-document sharing, groups and guests. People join by invite link and sign in with a password or your OpenID Connect provider.
  • Import and export. A Notion export, or a zipped Obsidian vault or folder of Markdown, becomes a workspace; a workspace exports as one .stuga.zip of Markdown, JSON Lines and its files.

[The built-in co-author's edits arriving as suggestions, accepted and rejected one by one]

Connect your agents

Agents connect over MCP. Settings → Your AI agents gives the setup for Claude Code, Claude Desktop, Codex, Antigravity, Cursor, VS Code, Kiro, Goose, LM Studio, DeepSeek Harness and Pi, with your node's address filled in, and the URL any other MCP client needs (docs/agents.md). Each agent brings its own model, so the node needs no AI key.

  • Scoped sign-in. Most clients sign in through the browser, where you tick the workspaces an app may use and choose Read only or Read and suggest changes. Revoke ends its access.
  • Narrow keys. A client without sign-in uses an API key, which can be confined to folders, made read-only or given an expiry.
  • Content only. Agents change documents and databases. Renaming, moving, deleting, sharing and the review setting stay with people.
  • Full audit. Every MCP call, reads included, is in the workspace's audit log. An event feed and signed webhooks report what changed.

Your data

Stuga sends no telemetry, and nothing reaches us unless a node admin turns on remote access. The node makes outbound requests only for features in use, such as an AI provider, and asks GitHub once a day for the list of releases unless a node admin turns that off. docs/privacy.md lists what the node sends and stores.

Install

  • A Mac with Apple silicon, on macOS 13 or later: open Stuga.pkg. It carries its own Postgres and Node.js (docs/install/macos.md).

  • Docker on x86-64 or arm64, such as a Linux server or a NAS: install.sh starts Postgres and the node with Compose (docs/install/docker.md):

    sh
    curl -fsSL https://github.com/stuga-dev/stuga/releases/latest/download/install.sh | bash

A node is one Node.js process and Postgres with pgvector and pg_search; people open it in a browser. It backs itself up every day by default and before every upgrade (docs/operations.md). docs/getting-started.md walks through the first hour.

Questions and feedback

Ask a question or share an idea in Discussions; report a bug as an issue.

Contributing

Issues and pull requests are welcome. CONTRIBUTING.md covers running Stuga from source, the tests and the conventions, and RELEASING.md how a release is cut. Contributions are accepted under a Contributor License Agreement (CLA.md); a bot asks on your first pull request. Report a vulnerability privately, as SECURITY.md describes.

License

AGPL-3.0-only, except integrations/, which is MIT; copyright notice in NOTICE. If you modify Stuga and offer it to others over a network, you share your changes under the same terms. The name and logo are not covered by the license: TRADEMARKS.md says how you may use them.

來源：README.md，提交 0c0a0b3

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v0.1.10最新Oct 2, 2026